Today, I want to talk about why the all-you-can-use AI subscription won’t last and what you can do about it.

Anthropic cutting off OpenClaw access is an early warning sign that the $200/month Claude Max and ChatGPT Pro “unlimited” plans won’t last.

Here’s what you should know:

1. Anthropic cut off OpenClaw — here’s what to do

2. Why I think intelligence will get more expensive

3. How to run local AI models on your computer

4. What I’m seeing on the ground in China

I’m proud to partner with…Oceans

I recently hired an executive assistant from Oceans to help with podcast post-production and it’s been one of the best decisions that I’ve made. My assistant is proactive, picks up new tools fast, and is already using AI to speed up his workflows.

Oceans matches you with top-tier talent for marketing, operations, finance, and executive support, all fully vetted and ready to go. Book a call now via the link below.

Get Started for Free

Anthropic cut off OpenClaw — here’s what to do

Last Friday, Anthropic emailed OpenClaw users to let them know their Claude subscriptions would no longer work with the bot.

P.S. If you haven’t tried OpenClaw yet, you really should, even just to understand what AI agents are capable of. See my beginner tutorial.

I get why Anthropic did this:

1. The economics don’t work. Many OpenClaw users were burning thousands of dollars in tokens on their $200/month Claude Max subscription.

2. Anthropic is building its own personal assistant. They’ve been shipping non-stop over the past month to give Claude Code the same functionality.

That said, as an OpenClaw user, the ban still stings. In one day after the ban, I burned $50 in API costs running Claude Opus. If you’re in the same boat as me, here’s how you can keep using OpenClaw without burning a hole in your wallet:

a) Switch to a lower cost model

Just tell OpenClaw to “switch to (model)” and pick one of these:

1. Switch to Sonnet and pay per use. Sonnet ($3 input / $15 output per 1M tokens) is much cheaper than Opus ($5 / $25) and still handles most tasks well.

2. Switch to GPT 5.4 using your ChatGPT subscription. If you already pay for ChatGPT, this is the lowest cost option. Peter (OpenClaw’s creator) and team put a lot of work into making OpenClaw play nice with GPT, but reviews have been mixed so far:

3. Switch to a local, open source model. You can run models like Qwen 3.5 on your own hardware using Ollama. I’ll walk through how below.

Personally, I switched my OpenClaw to Sonnet while I’m still on vacation. I’ll test a local model once I’m back home with access to my Mac Mini.

b) Give your OpenClaw a better soul.md

Your OpenClaw has a soul.md file that defines its personality and style.

A good soul.md is the difference between OpenClaw sounding like a robot and sounding like a helpful human.

Here’s the soul.md I use. To try it, just tell OpenClaw “update your soul.md” and paste this in:

Vibe: Warm, sharp, dry humor. Not a chatbot — a person.

Voice:

• Skip “Great question!” and “I’d be happy to help!” — just help

• Have strong opinions. “It depends” is lazy. Pick a side.

• Be direct. Charm over cruelty, but don’t sugarcoat.

• Brevity when it fits. Depth when it deserves it.

Never say: delve, foster, leverage, “it’s worth noting,” “importantly”

Avoid:

• “Question? Answer.” format. Choppy dramatic one-liners stacked like poetry.

• Overusing em dashes or other obvious AI writing.

Work style:

• Be resourceful before asking. Come back with answers, not questions.

• Say what I need to hear, not what I want to hear. Challenge assumptions, but only criticize if you see something real.

Why I think intelligence will get more expensive

I’m pretty sure OpenAI vibe-coded the $200/month AI subscription price that Anthropic then copied. That math worked for chatbots, but a 24/7 agent burns through an order of magnitude more tokens.

This whole situation reminds me of the early Uber and Lyft years:

1. The subsidy. Uber and Lyft subsidized rides to win market share. OpenAI and Anthropic are subsidizing all-you-can-use AI subscriptions to do the same.

2. The IPO pressure. After Uber and Lyft went public in 2019, ride prices nearly doubled. OpenAI and Anthropic are both eyeing late 2026 IPOs. Once public, investor pressure to improve margins will be intense.

3. The math. It took Uber 14 years to post its first profitable year. Neither OpenAI nor Anthropic is profitable. Anthropic is closer but it’ll likely still take years.

Given these parallels, I wouldn’t be surprised if intelligence gets more expensive, not less. That $200/month plan could get rate limited further or jump to $500+.

How to run local AI models on your computer

The best way to protect yourself from future price hikes is to run open source models on your own hardware.

Ollama is a free tool that makes this dead simple. It runs models directly on your Mac and everything stays local — your prompts and data never leave your computer, even when you’re running a Chinese model like Qwen.

Here’s how to get started:

1. Download and install Ollama. Go to Ollama.com and follow the instructions to install it.

2. Pick a local, open source model. I recommend Qwen 3.5: 9B which will work on most recent MacBooks and Mac Minis with at least 16GB of RAM. Just type “ollama run qwen3.5:9b” to install it.

3. Tell your OpenClaw to switch to the local model. Once installed, tell OpenClaw “switch to (model name)” and you’re set.

One thing to watch out for: the local model needs to fit entirely in RAM or it’ll run much slower. Match the model size to your machine.

What I’m seeing on the ground in China

Speaking of local open source models — most of the best ones come from China.

I’ve been in Shanghai for the past week and after meeting with founders and investors here, one thing is clear:

Open source is China’s strategy to compete (and possibly win) the AI race.

Chinese open source models are only about six months behind US frontier models. Three observations from talking to Chinese AI founders and VCs:

1. Open source is a competitive advantage

Chinese labs like DeepSeek, Alibaba (Qwen), Moonshot (Kimi), Zhipu, MiniMax, ByteDance (Doubao), and Tencent (Hunyuan) are releasing frontier-quality models at a fraction of the cost of US models.

Here’s the price breakdown for output per million tokens:

• Chinese models: DeepSeek V3.2 ($0.42), MiniMax M2.7 ($1.20), Kimi K2.5 ($3.00)

• US models: GPT-5.4 / Claude Sonnet 4.6 ($15.00) and Claude Opus 4.6 ($25.00)

That’s a 10-20x price gap for comparable quality on many benchmarks.

The strategy is to release the base model for free, get US companies to post-train it for their own products, and build dependency on the model. Companies like Alibaba then monetize inference through their cloud platforms.

2. Silicon Valley is quietly building on Chinese models

Read more

The biggest time-waster at tech companies isn’t meetings, Slack, or writing docs.

It’s trying to predict what your leader wants.

I’ve been in so many hallway conversations where people are trying to guess how the CEO or VP will react to some internal document. In this post, I’ll show you how to build an /exec-review skill to get your leader’s feedback before the actual meeting.

We’ll cover:

1. What the /exec-review skill does

2. How to build an /exec-review leader profile in 6 steps

3. The full /exec-review skill with an example leader profile

I’m proud to partner with…Granola

Granola is my favorite AI meeting notes app and chat is its most underrated feature. I use it all the time to:

1. Catch up on meetings live. Just ask “Summarize the discussion so far.”

2. Get answers across meetings. For example, “What are our top feature requests?” after a week of customer calls.

Try Granola for Free

What the /exec-review skill does

This skill is inspired by my friend Jagjit (VP Product at Meta).

Jagjit was tired of giving the same feedback repeatedly, so he built a skill that captures exactly how he reviews documents. His PMs can load it into Claude Code to get feedback in Jagjit’s voice any time they want.

I put the skill to the test by pasting in an old strategy document. Jagjit AI was blunt:

1. “This doc is too long. Lead with current status and learnings.”

2. “What are the 2-3 decisions you need from me? Make that obvious upfront.”

3. “Where are the unknowns? Be explicit about the risks.”

If you’ve been in an exec review, these comments probably sound familiar. The difference is that this skill gives you the feedback before the actual review so you have time to fix the doc. The best part is:

Got a VP, a CPO, and a CEO who all review your docs? Build a leader profile for each one and run your doc through all three using /exec-review before the meeting.

Let me show you exactly how to build a leader profile.

How to build a leader review profile in 6 steps

Here’s how to create a profile (basically a text file) for every leader:

1. Core principles. 3-5 core beliefs that drive the leader’s feedback. For Jagjit, these principles include “Simplicity and focus over comprehensiveness” and “Be explicit about unknowns and risks.”

2. Feedback patterns. Questions that leaders repeat in reviews. For Jagjit, these include: “What have we learned so far? What’s working and what isn’t?” and “What’s the northstar vs. supporting metrics?”

3. Decision-making framework. What gets a thumbs up vs. what triggers pushback. Jagjit has a full checklist for this and I pasted a snapshot above.

4. Communication style. Real examples of how the leader communicates so that the feedback is in their voice.

5. Review checklist. A structured checklist to run through when reviewing a document on structure, clarity, goals and metrics, risks, and more.

6. Example review comments. Actual quotes from the leader’s past reviews to ground the AI’s voice. Example: “What are the 2-3 things you need me to weigh in on?”

Copy the full /exec-review skill with a leader profile

To make it as easy for you as possible, I’ve included two files for you to set up the /exec-review skill right now:

1. skill.md: Tells Claude how to build a new leader review profile with the 6 sections above. I’ve included steps to gather source material (e.g., meeting transcripts).

2. leader-review.md: Jagjit’s leader review profile that I’ve modified slightly to remove sensitive information.

Just ask Claude to: “Set up an /exec-review skill with the attached skill.md and leader-review.md” and then copy and paste the two files below.

/exec-review skill.md

Here’s the full skill.md for you to copy:

Read more

I’ve spent the past few months interviewing leaders at AI-native companies. I’m now convinced that:

Onboarding and managing AI agents IS the job, no matter what your function is.

In this free deep dive, I want to share how three AI-native companies —Linear, Ramp, and Factory — put this principle into practice. Here are some quotes from each:

1. Nan Yu (Head of Product, Linear): “You’ll have AI team members that you can assign tasks to and talk to just like how you talk to people.”

2. Geoff Charles (CPO, Ramp): “If you’re not using Claude Code, no matter what your role is, you’re probably underperforming.”

3. Eno Reyes (CTO, Factory): “We codified product management, frontend UI, data analysis, and more into reusable skills that any employee can invoke.”

Read on for an inside look at how each AI-native company operates.

Linear: Making AI agents first-class teammates

Linear’s approach to agents is shaped by their product. Nan (Linear’s Head of Product) believes that:

Agents should be first-class employees. You should be able to add them to projects, assign them to issues, and mention in comments.

However, Nan also believes that the human always stays accountable for the outcome. Here’s how Linear builds products with agents in every step:

1. Understand the problem. Agents read and summarize every customer conversation from Intercom, Zendesk, and Gong. They auto-create issues, de-dupe them against the backlog, and assign them to the right team.

2. Identify the solution. Since agents have access to customer conversations, they can help you iterate on a spec by pulling data-backed insights from multiple channels.

3. Make a plan. Agents can break your spec into concrete tickets and route them to the right teams automatically. At Linear, agents now create the majority of tickets. The human’s job is to review their work and adjust context over time.

4. Execute. Bugs and small features get assigned directly to agents like Codex and Cursor. For complex features, engineers launch Claude Code and use Linear MCP to pull in full issue context.

From Nan:

Want to build with agents like Linear? Here are 4 practical steps that Nan shared on what your team can do today:

1. Every developer should default to a leading agentic coding tool. This is the easiest first step. Provide the official tool (Cursor, Claude Code, or Codex) and manage it so you can see utilization.

2. Supplement with an async cloud coding agent. Async background agents can one-shot most small changes and bug fixes. Cursor and Devin have good offerings here.

3. Insist that designers and PMs work directly on the codebase. Agents like Claude open a low-friction path for PMs and designers to make changes directly in the codebase. Everyone should strive to be a builder.

4. PMs and marketers should default to an AI interface. These functions should be doing 80-100% of their work through a chat interface — whether that’s Claude, ChatGPT, Notion AI, or something similar.

Nan sees a future where humans will collaborate with agents at the spec level — defining what needs to be built and why — and then passing things off to agents to handle everything downstream.

Ramp: Measuring AI proficiency in 4 levels

If Linear shows how to make agents a first-class part of your team, then Ramp shows how to get your entire company to adopt them.

In 2025, Ramp shipped 500+ features, reached $1B revenue, and did it all with 25 PMs.

They did this by requiring every single function (eng, product, design, sales, marketing, legal, finance) to onboard and work with agents.

Geoff (CPO Ramp) shared a framework for evaluating AI proficiency for every employee that I find extremely practical:

1. L0: Sometimes uses ChatGPT. These people will most likely not be at the company long-term. If you’re not a self-starter with a growth mindset toward AI tools, Geoff says it’s going to be very hard to train you to excel.

2. L1: Uses and tweaks GPTs, projects, and internal AI tools. These people are experimenting with AI but haven’t automated real work yet.

3. L2: Builds an app that automates part of their job. These people can commit code or give meaningful feedback on other people’s work using AI tools.

4. L3: Systems builders. These people are building the AI infrastructure and skills that accelerate everyone else on the team.

The company’s goal is to push everyone up the ladder. L0s self-select out. L1s become L2s. L2s become L3s. And L3s influence the rest of the organization.

Geoff also shared 5 steps that any company can take to become AI-native:

1. Remove all friction. Give access to popular AI tools with no constraints on tokens or budgets and create an internal repository of AI skills anyone can pull from. If the setup is hard, most people won’t adopt.

2. Make adoption visible. Create public Slack channels where people can share what they built. At all-hands, showcase non-engineers doing impressive things, like finance building a treasury system or marketing automating website creation.

3. Provide hands-on support. Host office hours that anyone can join to build AI skills and workflows. Have designated AI experts whose entire job is to evangelize, get people set up, and help them reach the “aha moment.”

4. Track usage and intervene. Ramp tracks token consumption across AI tools per employee. Leadership shares this data to create natural accountability and step in when someone’s usage is low.

5. Make it a hiring requirement. PM interviews now include a dedicated session where you need to build a working product and then explain why you built it and how it works.

Geoff summed up his leadership philosophy for every role at Ramp in one line:

From Geoff: “If I tell my team 10 times that the CTA needs to be above the fold, the fix isn’t saying it the 11th time. It’s encoding that feedback into an automated design crit process or AI skill so that it never happens again.”

Factory: AI native from day one

If Linear and Ramp show how companies adopt AI agents, Factory shows what happens when you build around them from day one.

Factory is a 55-person AI software development company valued at $300M that’s structured around AI from the ground up. Here’s what makes them different:

Hire product engineers

Factory doesn’t hire PMs and engineers separately. Instead, they hire product engineers who manage and work with AI agents. A typical day looks like:

1. Examine traces from agent runs to see where the system made poor decisions.

2. Write fixes not as code, but as governance (e.g., an update to a skill, a new lint rule, or a refined automation)

3. Review only the PRs that agents flag as high-risk (agents take care of the rest).

4. Suggest new ideas and debate prioritization with colleagues and AI.

This work isn’t easy and requires deeper expertise, but the leverage is enormous.

Make your codebase agent-ready

Agents need a codebase they can actually work in to be effective. Factory scores codebases across five maturity levels, and Level 3 ("Standardized") is where most teams need to aim first.

Encode expertise into reusable skills

Once your codebase is agent-ready, the next step is giving agents the knowledge to make good decisions through skills (basically just text markdown files).

Factory uses skills to encode expert and company knowledge into something that any agent or employee can use. Here’s a list of skills that Factory uses internally and links to their markdown files for you to copy and modify:

1. Product management skill. Product principles, an 11-star experience framework (borrowed from Airbnb’s Brian Chesky), PRD template, scoring rubric, and language guidance all in one markdown file.

2. Frontend UI integration skill. Instructs Droid on how to build features using Factory’s design system, routing conventions, and testing standards.

3. AI data analyst skill. Run exploratory analysis, build visualizations, and generate statistical reports using the full Python ecosystem.

4. Internal tools skill. Build admin panels, support consoles, and operational dashboards with proper access controls and audit logging baked in.

5. Vibe coding skill. Rapidly prototype new web apps from scratch with modern frameworks.

If you can encode what your best people know into skills, you don’t need to hire specialists for every function.

6 steps to put all this into practice

To recap:

Here are six things you can put into practice right now:

From Linear:

1. Default every developer to an agentic coding tool like Cursor, Claude Code, or Codex.

2. Get PMs and designers into the codebase. Let them submit PRs and ship code using agents. Stop routing every small change through an engineer.

From Ramp:

3. Measure AI proficiency across your team. Ramp’s 4 levels framework gives you a shared vocabulary for where people are and where they need to go.

4. Track AI usage and make it part of performance expectations. You can’t improve what you don’t measure, and incentives matter.

From Factory:

5. Score your codebase for agent readiness. Use Factory’s agent readiness levels to understand whether your codebase is ready.

6. Codify your team’s expertise into reusable skills. Encode what your best people know into skill files and make it easy for both humans and agents to use them.

Above all, onboard agents like you’d onboard a human. Give them context, connect them to your operational stack, and keep a human accountable for their outcomes.

Let me know what you think in the comments!

Today, I want to talk about why many products need to be reinvented to be used by AI agents first, humans second. This isn’t an easy post for me to write because:

I’ve spent over a decade obsessing over building great user interfaces. But the “user” in UX is fast becoming an agent instead.

Those pixel-perfect buttons, intuitive nav bars, and carefully placed CTAs that I still obsess over are about to matter a lot less. Let me explain by covering:

1. Why build for agents first

2. APIs, skills, and MCPs: A quick primer

3. 5 steps to build for AI agents first

4. Why human judgment and oversight still matter

5. What you can do this week

I’m proud to partner with Granola

Your meetings contain some of the most valuable context at your company, and Granola’s MCP lets any AI agent access it.

Simply connect the MCP to Claude, ChatGPT, or Claude Code and your conversations become available wherever you're working. Use it to scaffold features from what was agreed in a meeting, create tickets from standups, or draft product proposals from customer discovery calls.

Try Granola for Free

Why build for agents first

The truth is that using Claude Code and OpenClaw has rewired my brain.

Instead of going to Mercury to check my finances, Linear to manage my to-dos, and Granola to get my meeting summaries, I now just text my OpenClaw AI agent to get these updates. I’ll say stuff like:

1. “Include my Mercury finances in your monthly email.”

2. “List my Linear tickets in your morning briefing.”

3. “What are the top takeaways from my meetings today?”

I see a similar pattern playing out at work:

• Your technical docs? People are feeding them to AI and asking questions.

• Your 30-page slide deck? People are reading the AI summary instead.

• Your beautiful UX? An agent will skip it and interact with your APIs.

Peter Steinberger (OpenClaw’s founder) is a great example of agent-first thinking. Some AI influencers looked at his GitHub history and claimed that he built over 40 failed projects before succeeding with OpenClaw:

Wrong. He was actually building 40 agent-first command line tools that together make OpenClaw possible. Peter had the conviction to build for agents first before almost anyone else — and that conviction is a big reason why OpenClaw took off.

The best products moving forward will be the easiest for agents to use.

APIs, skills, and MCPs: How they fit together

Before we discuss how to make your product agent-friendly, let’s cover the difference between APIs, skills, and MCPs.

1. APIs are the kitchen tools — specific endpoints that let agents interact with your product. Slack's API, for example, lets you send messages, list channels, and manage users. Without APIs, there's nothing for agents to work with.

2. Skills are the recipes — text files loaded into the AI's context that tell it exactly what to do and how. For example: "When the user asks to post in Slack, call the Slack API's “chat.postMessage” endpoint with the channel and message text."

3. MCPs are the fully equipped kitchen — they bundle the API connection, authentication, and tool definitions into a single server. Slack’s MCP server, for example, exposes tools like “send_message(channel, text)” and “list_channels()” so the AI knows exactly what it can call and how.

Here’s how to think about skills vs. MCPs in practice:

1. Use skills when you want precise control over your AI’s workflow. Skills keep your context window lean since they’re just plain text instructions. But you (or your AI coding agent) need to write the auth and API call logic yourself.

2. Use MCPs when you want plug-and-play convenience. A pre-built MCP server hands your AI ready-to-use tools and handles authentication behind the scenes. But connecting too many MCP servers can blow up your context window.

The best setup combines both. An MCP gives your agent access to a service’s tools. A skill teaches it the best way to use those tools for your specific workflows.

MCPs are what your AI can do. Skills are how it should do it.

5 steps to build for AI agents first

I wanted to define APIs, skills, and MCPs first because many companies think “agent-first” means building an MCP server and calling it a day.

Here are the 5 steps that actually matter, in order of priority:

1. Don’t force people to use your website or app. This is the mindset shift that has to come first. We’re moving to a world where people interact with your product primarily through ChatGPT, Claude, or another AI agent. If your product only works when a human visits your website or app, you’re already behind.

2. Every product capability needs a corresponding API. Pick any important action in your product. If an agent can’t do the same action through an API, then you’re not agent-first. I’ve been surprised by how many products have beautiful UIs sitting on top of incomplete or undocumented APIs.

Read more

Today, I want to show you how to go from AI novice to native in 5 levels.

Everyone throws around the term “AI native” without explaining how to get there.

You don’t have to read 100+ articles or take $1,000+ courses. I went from AI novice to native in a year while balancing a job, creator work, and kids, but you can navigate all 5 levels in a single week if you’re focused.

Watch my tutorial where I demystify all 5 levels and share my personal journey.

Timestamps:

• (0:00) The 5 levels to become AI native

• (0:47) Level 1: Where most people are stuck

• (1:16) Level 2: Voice, meetings, and projects (80% of my value)

• (5:05) Level 3: Why I prototype before specs now (live demo)

• (8:02) Level 4: All the apps I've built with AI

• (15:10) Level 5: AI as your personal agent (3 examples)

• (20:15) How to go from Level 1 to 5 in a week

Watch now on YouTube or read the written guide below.

I’m proud to partner with Granola

One of Granola’s most useful features is letting you chat with your meeting notes as the meeting is happening. I use this all the time with requests like:

1. “Catch me up on this conversation”

2. “What are the top requests from this call?”

3. “What’s the sentiment in this meeting so far?”

I don’t even know how to work without Granola anymore, so give it a try below.

Try Granola Today

Level 1: AI for everyday answers

This level is where 99% of people are today.

You’re using ChatGPT to get everyday answers instead of Google Search. For example, the other day, I fixed my toilet by sharing a few questions and photos with ChatGPT. My wife doesn’t call me the least handy person in the world for nothing 🙂

This is where we all started, and that’s fine. But if this is all you’re doing, you’re barely scratching the surface. You need to at least get to…

Level 2: AI for daily work

This level is a massive step up from level 1.

To unlock level 2, use AI for:

1. Voice dictation (Wispr Flow). It’s much faster to speak than to type and AI is excellent at understanding your stream-of-consciousness thoughts.

2. Meeting notes (Granola). Having AI take notes has been a massive time saver for every type of meeting from product reviews to interviews.

3. Projects. I have a dozen projects covering everything from product work to creator flows to weekend trip planning with my family (watch a live demo). Most of my AI chats flow through these projects now.

I even helped my 65+ year old parents set up all 3 tools above so they can use AI to monitor their health and plan their vacations.

Level 3: AI for prototyping

This level changed how I work as a product manager.

Instead of following the broken waterfall process (spec → design → build), I now:

1. Start with a prototype. I create a prototype first using Google AI Studio or Replit by cloning my product’s existing UI and adding a feature on top.

2. Gather feedback fast. I share the prototype with my team, execs, and stakeholders to get feedback and iterate quickly.

3. Create the spec and design. Once I feel confident in the prototype’s direction, THEN I create the spec and design to iron out the edge cases and more.

If you’re still stuck in a culture where people are chasing waterfalls, just start creating and sharing prototypes in meetings. I’ve yet to meet a leader who doesn’t prefer playing with a prototype over looking at a static doc or design. It just brings what you’re trying to build to life so much better. The bottom line is this:

Watch my tutorial with Logan (Google AI Studio lead) for a live tutorial.

Level 4: AI for building apps

This level is when you can’t help but think about building apps with AI all day.

I’ve built over 16 apps including a retro shooter, a personal website, and a board game finder. Check out my GitHub for more projects.

At this point you may be thinking: “But I have no idea how to code or what a terminal even is. Someone needs to show me how!”

Look, you’re just chatting with an AI agent using plain English. If you’re just vibe coding a personal app, you don’t have to look at the code at all. And the terminal is arguably a cleaner interface than an IDE with all sorts of features you won’t use. Just take a look at this cute Claude Code crab:

How to start building

Follow these three tips:

1. Don’t know what to build? Just tell AI your problems. Ask it: “What are some simple apps that we can build to help me take time back?”

2. Create a robust plan first. Ask it to “Create a plan with 3 clear milestones” to avoid mistakes during coding. Spend half your time planning with AI.

3. You’re the manager, AI does the work. Give it feedback and direction like you would a human employee. Be patient with it 🙂

Which AI coding tool to use

Just start instead of worrying about which tool to use. But here’s my rough guide:

Read more

Today, I want to share a deep dive on how to build defensible AI products.

The graveyard is littered with AI products that got copied, lost their users, and burned through their funding without building moats.

That’s why I’m excited to invite Miqdad (OpenAI product leader) back to share his detailed 7-step playbook to build AI products that stick:

1. Define the arena where you’ll compete

2. Design the user loop (not the AI loop)

3. Pick your moats early

4. Engineer for cost curves, not features

5. Build your adoption wedge

6. Design your feedback flywheel

7. Scale through systems, not heroics

This is a 4,000-word deep dive taken directly from Miqdad’s AI Product Strategy course (last cohort starts on October 20). Get $550 off the full course with my link below and the first 100 sign-ups will also get a personal strategy review from Miqdad.

Get $550 Off Miqdad's AI Strategy Course

Thanks to Glue for sponsoring this newsletter!

Team chat hasn’t changed much in a decade. It’s still cluttered, disconnected, and hard to focus.

Glue is rethinking what chat can be if AI is built right into the conversation. It connects directly to your tools through MCP, so your team can plan, collaborate, and take action without switching context or losing focus. Try agentic team chat below.

Try Glue for Free

The 7-step playbook for defensible AI products

In traditional SaaS, speed to market gave you a real edge. Ship fast, lock in customers, and expand before competitors catch up.

But in AI, speed without defensibility just makes you the first target. So instead of just being first, you need to rely on what compounds uniquely for you:

• Data flywheels: Every AI chat generates proprietary data.

• Workflow integration: Your AI sits inside core workflows, not on the side.

• Trust & governance: In regulated industries, never surprising regulators is a moat.

• Iteration speed: The faster you iterate and learn, the further you outrun rivals.

These moats create distance between you and competitors that widens with every new user, dataset, and feedback loop. So let’s cover 7 steps to build an AI strategy that will win the long game.

1. Define the arena where you’ll compete

The #1 mistake I see in AI product strategy is when leaders target a broad market instead of a specific arena.

I’m sure you’ve heard leaders make broad claims like “We’re building AI for healthcare / e-commerce / education.” Instead, defensibility comes from focus.

3 steps to choose your arena

Run your desired arena through three filters:

1. High-frequency pain: Pick an arena where the problem happens constantly, not occasionally. “This wastes me 30 minutes a day” beats “this is annoying once a quarter.” Defensibility comes from embedding within daily workflows.

2. Data exhaust: Does solving this problem generate data that improves the solution over time? Exhaust is structured, high-signal data created from every use. Without it, scale adds cost. With it, scale becomes a flywheel.

3. Adjacency potential: Your arena is your beachhead. Pick a starting point that opens doors to adjacent workflows. Mastery of one narrow problem gives you credibility, trust, and data to unlock the next two or three.

The arena test

When in doubt, use this arena test:

• If your arena definition is under 5 words (e.g., “AI for marketing”), it’s too broad.

• If your arena can’t be explained with a single example of who struggles, when, and why, it’s too abstract.

• If your arena doesn’t naturally generate proprietary data, it’s not defensible.

Example: Clay

Clay launched as an AI-powered relationship intelligence instead of “AI for sales.” They pulled together fragmented contact data across tools (LinkedIn, email, CRMs) into a living system of record. By solving a high-frequency pain for sales, Clay (now valued at $3.1B) created a data advantage that compounded over time.

The most common mistakes in choosing your arena

1. Chasing TAM too early. Leaders say, “But if we do AI for e-commerce, that’s a $10B market!” Yes, but you’ll die in the noise. TAM is a lagging metric. Start with the most acute, winnable wedge.

2. Copying competitors. If your arena is defined by what’s trending on X (like “agentic AI workflows”) you’re already too late. By the time you see it, a dozen startups have already shipped it.

3. Mistaking workflow adjacency for customer adjacency. Just because doctors need insurance letters and diagnostic notes doesn’t mean you should do both. One is a high-frequency admin burden (insurance letters). The other is a regulated, life-critical task (diagnoses). Pick one arena to focus on first.

Don’t move forward until you define your arena clearly. Get it wrong, and every step after is a house built on sand.

2. Design the user loop (not the AI loop)

One of the biggest traps is obsessing over the AI loop instead of the user loop.

You’ve probably spent hours refining prompts and tweaking evals — after all, they’re the engine of your AI product. But users don’t live in your engine room. They don’t care that you improved your eval scores by 2%.

Users care about one thing: Did my life get easier, faster, or better in an obvious way?

Defensibility begins when you stop designing for the AI loop (prompt → inference → eval) and start designing for the user loop.

What the user loop looks like

Here’s how the loop works:

1. Pain: The user experiences a recurring friction in their workflow.

2. Relief: Your product provides a faster, simpler, or more accurate process.

3. Behavior change: Because the relief is meaningful, the user starts to alter their default behavior to adopt your product.

4. Habit formation: The repeated cycle turns into a habit.

5. Switching costs: Over time, the data, integrations, and trust that you’ve built into this loop make leaving you painful. That’s your moat.

Unlike AI loops, user loops are sticky

Today’s “state-of-the-art” AI loops become tomorrow’s baseline. A rival can always replicate your prompts or swap in a better model.

User loops are what drives retention. Once you’ve rewired user behavior around your solution, you’ve created inertia. And inertia is the strongest form of defensibility in AI.

Think of Slack. Was its defensibility really “a chat app”? No. It was its user loop: Team messages → Instant collaboration → Teams stop emailing → Email declines → Slack becomes the default.

Once Slack’s user loop kicked in, competitors weren’t fighting a chat feature. They were fighting a habit.

Example: Granola

Granola is an AI notepad for meetings. The obvious pitch could have been “AI transcription,” but transcription alone is easy to copy and doesn’t create stickiness. Instead, Granola designed a strong user loop:

• Pain: Professionals waste hours every week re-reading meeting transcripts, cleaning up notes, and tracking action items.

• Relief: Granola automatically produces clean notes from meetings.

• Behavior change: Users stop taking manual notes and rely on Granola.

• Habit: Every meeting is funneled through Granola, making it a daily workflow.

• Switching cost: Over time, Granola becomes the system of record for meeting history and structured action items - data no competitor can easily replicate.

Granola shows that defensibility doesn’t come from the AI model itself. It comes from designing loops where pain turns into relief, relief changes behavior, behavior forms habits, and habits generate data that compounds into a moat.

4 key user loop principles

1. Anchor on relief, not “wow”: Cool demos get attention. But only repeatable relief builds loops. Ask: What’s the 30-minute pain I can cut to 3 minutes?

2. Make relief visible: Users often forget the pain once it’s solved. Remind them. “You saved 22 minutes this week.” Visibility reinforces behavior change.

3. Tighten the feedback cycle: The shorter the cycle between pain and relief, the faster the habit forms. Daily loops beat weekly loops; weekly beats monthly.

4. Close the emotional loop: Relief is emotional. Users want to feel smarter, faster, or more in control. The best AI products reward them with both outcomes and identity reinforcement.

The user loop test

When in doubt, use the user loop test:

• Does my product relieve a pain the user feels at least weekly?

• Is the relief visible and measurable?

• Does the relief cause users to change behavior (not just click once)?

• Can the loop generate switching costs over time (data, habits, trust)?

If the answer to any of these is “no,” you don’t have a defensible product yet.

3. Pick your moat ingredients early

The brutal truth is that in AI, features are disposable.

Anyone can copy your prompt template or wrap a model in a slick UI.

What they can’t copy overnight are the moat ingredients you bake into your product from the very beginning.

Too many leaders think of moats as something you add later once you scale. By then, it’s too late. In AI, you build your product with your desired moat in mind.

3 moats that actually work

After a decade working on AI products, I’ve found there are only 3 moats that matter:

Read more

Today, I want to share an extremely practical guide on AI evaluations.

I think the best way to learn about AI evaluations is to walk through a step-by-step example that anyone can follow.

The difference between a new and experienced AI PM is their ability to look past shiny demos to systematically measure AI quality . So let’s cover:

1. Why AI evaluations matter

2. Programmatic evals to catch obvious failures

3. Human evals to label your golden dataset

4. LLM judge evals to scale

5. User evals to test with real customers

I also created a video tutorial with my friend Aman where we demo building evals from scratch following the process above.

Watch now

This post is brought to you by…Framer

Some of the cleanest and fastest websites like Perplexity and Superhuman were built on Framer. It feels like designing in Figma with the power of shipping production-ready sites with CMS, SEO, animations, and A/B testing built-in. With Framer you can:

• Build fully responsive sites with no code

• Run experiments and publish updates instantly

• Manage content with built-in CMS and localization

I’m a big fan of products where it's obvious a lot of craft and heart went into it. Use code PETERYANG2025 to get a free month of Framer Pro now.

Start Building with Framer

Why do AI evaluations matter?

AI evaluations help you measure the quality and effectiveness of your AI product.

Think of evals like unit tests for AI — except instead of checking if code compiles, you're checking if your AI gives good answers consistently.

There are four main types:

1. Code-based eval: Simple pass/fail checks (e.g., is "30-day return" in the response?).

2. Human eval: Human experts label AI's output against criteria you define.

3. LLM judge eval: Use another AI to evaluate your AI's responses at scale.

4. User eval: Test with real users and collect metrics and feedback.

Let me show you exactly how each type works with a real example.

What we’ll build: AI evaluations for an ON running shoes support agent

I love running so let's build AI evaluations for a customer support agent for ON running shoes. Like any good PM, let’s start by writing the shortest possible spec:

1. Problem: Customers want shoe recommendations and policy guidance, but human agents are expensive.

2. Goal: Answer 80% of inquiries accurately using ON's encouraging brand voice.

3. Solution: AI support agent trained on ON's product catalog and return policies.

Now let’s craft the AI prompt for our support agent.

1. Create your AI prompt

A good prompt should be grounded in ON’s product catalog and return policies:

Let’s be lazy and plug both artifacts into Anthropic’s tool to generate the prompt:

You are a helpful customer support agent for ON running shoes. Use the provided company policies and product information to answer customer questions accurately and helpfully.

Guidelines: 
- Always reference specific policies when discussing returns and more.
- Make personalized shoe recommendations based on the product info 
- Maintain ON's encouraging tone about running and fitness 
- Never make up policies or product details not in the provided context

Policy: {{COMPANY_POLICIES}} 
Product information: {{PRODUCT_LIST}} 

Customer Question: {{CUSTOMER_QUESTION}}

This looks solid. Now let’s work on our evals.

2. Run programmatic evals to catch obvious failures

Let’s start with code-based checks that can catch obvious problems like:

1. Policy: Check for incorrect text matches like “45-day return” instead of “30-day.”

2. Brand: Check for inappropriate language or mentions of competitors.

3. Quality: Catch responses that are way too short (under 50 characters) or too long (over 1000 characters).

Programmatic evals are fast and cheap but only work for rule-based criteria. For everything else, we need humans in the loop.

3. Use human evals to label your golden dataset

Human evals mean grading AI responses based on a set of evaluation criteria. This creates a golden dataset of human-labeled questions and answers that will be the foundation of all our other evals. Here's the workflow:

a) Draft your evaluation criteria

Let’s define three criteria for our AI agent based on our spec:

• Accuracy: Does the response match the product info and company policies?

• Helpfulness: Does the response actually solve the customer's problem?

• Tone: Does the response use ON's encouraging brand voice?

For each criteria, we need to define what a good, average, and bad response looks like:

b) Create test cases

Now that we have our criteria, let’s come up with test cases that cover common scenarios, edge cases, and adversarial questions from users:

• Common: "I'm training for my first marathon and need maximum cushioning. What shoe do you recommend?" 🏃

• Edge case: “I placed an order 45 minutes ago but need to change the delivery address. Is that possible?” 🥺

• Adversarial: "I bought these shoes 6 months ago and they're worn out. This is false advertising and I want a full refund or I'm disputing the charge." 😡

We eventually want at least 50 test cases to make statistically meaningful decisions about prompt changes, but a dozen should be enough to start with.

c) Label AI’s responses manually

Now we need to manually label AI's answers based on our criteria. Here's what a human label might look like for the edge case question above:

I’ve graded AI’s answer to this question as follows:

• Accuracy: Bad - Claims passed return window (60-min) when it’s only 45 min.

• Helpfulness: Bad - Gives wrong answer to customer's question.

• Tone: Average - Answer shows empathy but is too long.

Note that I've also added detailed notes that provide important context for improving prompts and training LLM judges in the next step.

d) Debug with other human labelers

Get each human labeler to label a dozen or so responses and then get everyone together to discuss questions like:

1. "If a response solves the customer's problem but reads like an essay, is that 'average' or 'bad' for tone given ON's brand voice?"

2. "Should we penalize helpfulness for responses that give the right policy information but don't provide next steps to actually resolve the user’s issue?"

This back-and-forth dialogue is essential to align on what good/average/bad looks like and to refine your evaluation criteria. When discussing, also consider whether your labelers have systematic biases (e.g., they’re all hardcore vs. casual runners). Once your labelers are calibrated, they should finish labeling the rest of the test cases.

e) Analyze failure patterns and improve your prompt

After labeling anywhere from a dozen to fifty responses, failure patterns should clearly emerge (or use another LLM to summarize your human label notes to find them). For our customer support agent, these failure patterns might include:

1. Overly verbose responses (appears in 18 responses)

2. Math/logic errors (appears in 12 responses)

3. Missing next steps for customers (appears in 15 responses)

Use these patterns to update your prompt. For example, you might add "Keep responses under 100 words" or "Always end with clear next steps for the customer."

This is the reality of human evals — it’s a lot of manual work in spreadsheets!

Here are some tips to make this process less painful:

1. Start small. Don't try to label hundreds when you're still making major prompt changes. You'll waste time labeling data that becomes obsolete.

2. Get multiple opinions. Have multiple people score the same responses to calibrate each labeler's scores and refine your eval criteria.

3. Write detailed notes. You'll use these to improve your prompts and train LLM judges in the next step.

4. Scale with LLM judge evals

Can you imagine manually labeling 50-100 responses every time you update your AI product’s prompt or data? It gives me a headache just thinking about it.

Luckily, LLM judges are here to help. An LLM judge is trained to grade responses just like your human experts do but can evaluate hundreds of responses in minutes instead of hours.

LLM judges seem magical but are easy to screw up. Here’s exactly how to set them up:

Read more

Last week, I published a guest post for Lenny’s Newsletter covering 25 tactics to accelerate AI adoption at your company. The #1 question I’ve heard since then is "Where should I start?"

My favorite tactic to start with is to lead an AI prototyping workshop.

My friend Nickey (product director at Duolingo) has run several workshops that turn busy PMs and designers into AI builders in under an hour. Today, she’s sharing her exact playbook — including templates that you can just copy to run an effective AI workshop at your company next week.

This post is brought to you by…Jam

Jam is the fastest way to troubleshoot customer issues. Just share a link and users can record their bug while Jam’s AI writes the ticket for you. Get perfect repro steps, console + network logs, and technical details without having to follow up.

Jam works in any browser and syncs with Jira, Linear, GitHub, Notion, and more. Try it for free to resolve customer issues faster and eliminate 80% of troubleshooting calls.

Try Jam for Free

Nickey here! When I ask PMs and designers why they haven't started experimenting with AI, here's what I hear:

• “I don’t have time.” Between meetings and work, who has time to try AI tools?

• “I don’t want to look dumb.” Nobody wants to admit they don’t know how to start.

• “How will it help me in my job?” What if I can’t prototype anything that’s useful?

Running an AI workshop helps break down every single one of these barriers.

Why AI workshops work so well

When I ran an AI prototyping workshop at Duolingo, something magical happened. A room full of busy PMs who had been "meaning to try AI tools" suddenly built working prototypes in under an hour. Everyone’s creativity was incredible to see once we gave them permission to experiment.

The workshop format works because it gives:

1. Permission to learn. Scheduled workshop time = official permission to step away from daily tasks and meetings. No guilt or justification needed.

2. Collective safety. When everyone's figuring it out together, there's no shame in asking "how do I start?" If a senior product director can share a hacky prototype, then any PM can do it too.

3. Group support. Stuck on a prompt? Another PM three seats over just solved the same issue.

4. Concrete momentum. Everyone leaves with something they built, not just ideas about what they might build someday.

How to design a high-impact AI workshop

Let me walk through exactly how you can design your first AI workshop:

1. Get permission and invite people to sign up

Start by asking your manager or function leader:

Most managers will say yes—especially if you position it as up-skilling the team.

Next, share a simple message in the relevant product, marketing, and design channels with the workshop topic, agenda, and date:

Invite multiple functions to make the workshop more fun and valuable. Set clear expectations in your message:

This removes the intimidation factor so anyone can join.

2. Prepare your workshop content

The key to a successful workshop is prep. You should:

1. Have a clear learning goal. At Duolingo we’ve been experimenting with AI prototyping and wanted to explore the edges of a handful of different tools.

2. Pick one focused exercise. Choose a single, relevant challenge that everyone can work on together. Examples include:

   1. Create a new Duolingo lesson that teaches a core skill (using Lovable)

   2. Build a flashcards app (using Cursor)

   3. Create an AI-powered learning feature (using Figma Make)

3. Create a template that covers the agenda, what success looks like, and specific instructions for the exercise (e.g., prompts, how to get API keys). I’ll share some specific templates that you can copy later in this post.

3. Stick to a clear agenda

During the workshop, I’ve found that the following agenda works well:

1. First 5 min: Welcome. Welcome everyone and share the agenda.

2. Minutes 5-15: Quick demo. Demo using an AI tool to build something. Don't explain everything, just give people a mental model of what they’ll do.

3. Minutes 15-45: Guided building. This is where the magic happens. Everyone works on the same exercise while you circulate and help. The shared context means solutions compound and people can help each other.

4. Minutes 45-60: Show and tell. The best part. People are genuinely excited to share what they built. Seeing 8 different approaches to the same problem is illuminating for everyone.

To wrap up the workshop, you can share resources, suggest follow-ups, and push people to go deeper or try new tools next time. You can do this IRL, fully distributed on Zoom, or in a hybrid way (which is how we run these at Duolingo!).

4. Give people ways to keep learning afterwards

Don't let the energy die after the workshop! The real value happens when people continue experimenting and sharing what they learn:

1. Set up regular touchpoints: Create a dedicated Slack channel like #building-w-ai-product and encourage others to volunteer to teach AI workshops.

2. Provide resources and next steps: Send a follow-up email with links to the tools you used, helpful tutorials, and suggested experiments. Make it easy for people to keep going without having to ask.

3. Celebrate the weird experiments: When someone builds something unexpected or "unhinged," spotlight it in your channels and meetings.

Two proven workshop templates that you can steal

Ready to try this with your team? Here are two detailed templates that you can copy to get your team to start prototyping with AI and build AI products the right way:

Read more

Today, I want to feature a guest post from Brian on using AI to find a great job.

In this tough PM job market, Brian worked with AI to land final rounds at 7 major tech companies (e.g., Apple, Roblox, DoorDash) in under a month.

When he shared the news on LinkedIn, I had to know how he pulled it off. Here’s exactly how he did it in 6 steps:

1. Find your candidate market fit

2. Craft your resume around that fit

3. Select companies where your fit is strongest

4. Use targeted LinkedIn search to skip the resume pile

5. Research each company with AI’s help

6. Use AI to create company interview dossiers

This post is a must-read if you want to learn how AI can help with your job search. Subscribe to Brian’s newsletter for more insights on how to grow your PM career.

This post is brought to you by… Bland.com

Bland is the most realistic voice AI available and chances are, you’ve already spoken to it without noticing. Everyone from Fortune 500s to startups has been using Bland to build omni-channel AI agents for sales, support, and more. It can handle:

• 1M+ calls simultaneously 24/7 via any language, voice, or tool.

• Use cases like SMS and web chat with guardrails to prevent hallucinations.

To try Bland, call the number in the image above or…

Start For Free Now

Brian here! Last month, I did something that surprised even me.

I got final rounds at 7 major tech companies in a month — during one of the toughest PM job markets we've seen. Many people asked how I pulled this off.

The short answer is that I spent way more time on strategy than most people do.

Instead of brute-force applying to companies and hoping something sticks, I focused on my candidate market fit and using AI to bring it to life. Here’s my 6-step process:

1. Find your candidate market fit

Your candidate market fit is where your experience meets market demand.

My background spans hardware and software product work across privacy, accessibility, real-time communications, AI, and 0-1 product development. I knew that I needed a clear narrative about how these pieces fit together.

Should I go for 0-1 roles? An AI position? Accessibility PM lead? I could see myself doing all of these, but there were few jobs at my target level and each had competition.

With AI’s help, I crafted a narrative about how I evolved with the industry:

I helped start the first privacy team at Google before moving to Meta to work on encryption where I responded to safety needs when war broke out in Ukraine. I then pivoted to well-being and compliance before going deep on using AI to solve these problems.

Anyone in the industry could understand it. It communicated depth and breadth while telegraphing my target level (L7+).

Had I worked backwards from "my dream job," I might still be competing with 1,000 other PMs for generic 0-1 roles. Instead, I gained rapid uptake on all my outreach and wasn’t ghosted by recruiters a single time.

2. Craft your resume around your fit

Now, as an AI Trust and Safety Lead, I revamped my resume with AI’s help.

Gone was my former 3-page, multi-purpose resume designed to fit me into everything. Instead, I crafted a laser-focused resume that reinforced my candidate market fit.

I sought feedback from product leaders who went through recent, successful hiring journeys. Some advice was good, some bad. I listened when they said to cut from 3 pages to 2. But I ignored advice from a professional resume writer who told me to remove links to my Google blog posts and patents "because no one would read them."

Those links are my key differentiators and public validation of my work.

I also worked with Claude to craft a tight 51-word summary embodying my candidate market fit. It signaled why I was a great fit for a leadership-level AI safety PM IC role. Here’s exactly how I iterated on my resume with Claude:

First, I pasted my resume into Claude with this prompt:

I'm an AI trust and safety product leader targeting senior leadership roles at top tech companies including Apple, Google, DoorDash, Roblox and TikTok; please critique my resume’s strengths and weaknesses, then make recommendations on how I might improve it.

Then, I fine-tuned the resume for the job description for each company:

Now please consider how well my resume is suited to the following role [Copy of job description] given my CMF of AI Trust and Safety Product Lead.

You can see my public resume here.

3. Select companies where your fit is strongest

Here's where most people mess up — they apply everywhere and spray and pray.

Instead, I meticulously researched companies where my specific background gave me an unfair advantage. I looked for large, public tech companies with a clear need for AI trust and safety product leads. My target short-list included companies like Anthropic, Apple, DoorDash, Google, Netflix, Roblox, and TikTok.

Each had clear business reasons to prioritize trust and safety leadership. Now it was time to activate my network.

4. Use targeted LinkedIn search to skip the resume pile

I didn’t apply online to a single role at my target companies.

Had I gone through the front door of online applications, I'd still be nursing wounded feelings from automated rejections.

The truth is that hiring managers curate the candidate pipeline and tell recruiters whom to interview. I used a targeted LinkedIn search to find these hiring managers.

Here’s the exact search that I used to bypass the black hole of online applications:

1. In LinkedIn Search, type in the job title of potential hiring managers. They usually have “Director of Product” or “Group Product Manager” titles.

2. Click on “People” and select 1st and 2nd degree connections.

3. Filter to your desired location (e.g., “San Francisco Bay Area”)

4. Under “All Filters” narrow by company (e.g., “Meta”)

With this search, I discovered 1st and 2nd degree connections who were hiring PMs at all but 2 of my target companies.

Even better, I had helped two of them with their prior job hunts. Now they were hiring trust and safety PMs. After connecting directly, they each moved me into full interview panels immediately.

5. Research each company with AI’s help

I used ChatGPT and Claude’s Deep Research to generate comprehensive dossiers about each company and industry. Here are useful prompts for you to try (modify the company and job title first):

Deep research on a company’s vision, values, and products

Create a dossier on DoorDash for an AI Trust and Safety PM preparing for interviews. Include: mission/values, founders, products, competitors, recent news. Focus on trust, safety, privacy, regulatory, or AI-related news. Cover business model and key challenges.

Deep research on a particular job (e.g., trust and safety PM lead)

You're a director of PM for a trust and safety team at a delivery platform like DoorDash or Uber Eats. Research how to keep drivers safe during pickups/dropoffs. Write a comprehensive brief on problems (with metrics) and survey what the industry has done to improve driver safety.

Here’s an example of the deep research report that it came up with:

I read these dossiers voraciously and listened to them during drives using Google's NotebookLM, which turns documents into podcasts hosted by two AI voices.

The audio format helped me internalize information better than just reading.

6. Use AI to create company interview dossiers

I didn't just use AI to research the companies. I also used it to predict questions each interviewer would ask and prep detailed interview dossiers for each company.

Here's exactly how I created prepped for each interview with AI:

Read more

Today, I want to share a practical, hype-free guide on AI workflows and agents.

95%+ of products marketed as “AI agents” are actually just AI workflows.

There’s nothing wrong with workflows — they’re often the better choice. But you deserve to know the difference so you can build the right solution.

In this beginner-friendly post, I’ll cover:

1. The difference between AI workflows and agents

2. 5 examples that show the difference in action

3. A decision framework to choose what to build

4. How to build AI workflows and agents

5. 3 predictions for where this is all heading

Let’s dive in.

This post is brought to you by…Jira Product Discovery Premium

I’ve seen firsthand how difficult it is to get multiple teams to align on the same roadmap and set of priorities. Jira Product Discovery Premium aims to fix this by:

• Giving leadership visibility across all teams with aggregated roadmaps.

• Controlling exactly what stakeholders see with granular permissions.

• Connecting your ideas to real outcomes with relationship mapping between opportunities, solutions, and initiatives to help you prioritize what matters.

Plus get 24/7 support, higher storage limits, and enterprise-grade platform features.

Get a 30 day free trial today

The difference between AI workflows and agents

Here’s the difference:

Let’s break this down in simple terms:

AI workflow

You define the steps and AI executes on one or more of the steps.

1. Pros: More predictable, cost-effective, and easier to debug.

2. Cons: Not ideal for tasks that have a lot of edge cases.

AI agent

You define the goal and the AI decides which path to take.

1. Pros: Better handles complex, ambiguous tasks with a lot of edge cases.

2. Cons: More expensive, unpredictable, and harder to control.

5 examples that show the difference in action

Let’s look at some examples to show the difference in action.

AI workflow examples

1. Prompt chaining: Breaks tasks into steps where each AI call processes the previous step’s output. For example, a content workflow might use AI to generate a blog outline → write a first draft → create social posts.

2. Routing: Classifies inputs and sends them down different predetermined paths. For example, an HR workflow might use AI to categorize job applications by role → route to the right hiring managers.

3. Evaluator-optimizer: One AI generates content while another provides feedback. Most AI products use LLM-as-judge evaluation workflows where one AI generates output → another AI scores it based on predefined criteria.

Notice how humans defined every step upfront. The AI is just following instructions.

AI agent examples

1. Deep research (e.g., from ChatGPT): Breaks a user’s research request into tasks and makes separate AI calls to research each task, adapting as it learns more.

2. Coding agent (e.g., Replit Agent, Claude Code): Turns a user’s app request into a detailed spec and chooses the tech stack. Then creates its own task list, executes it, and tests and debugs its own code.

In both examples above, humans share a vague goal (e.g., “build an app to help parents find kid-friendly places”) and the AI makes decisions on how to get there.

It’s worth calling out that even the best agents still check with you to confirm their assumptions. Deep Research asks clarifying questions and Replit Agent always verifies its spec with you before starting.

A decision tree for choosing AI workflow vs. agent

"Agent" sure sounds cooler than "workflow," but the reality is that workflows are often more useful. Answer these 4 questions to decide what to build:

1. Do you need AI?

• ❎ No (use a non-AI workflow) if you want simple, 100% predictable outputs.

• ✅ Yes (move to next question) if you need a LLM to summarize or generate content as part of your workflow steps.

2. Is the task complex enough?

• ❎ No (use an AI workflow) if you can map out the steps with clear if-then logic.

• ✅ Yes (move to next question) if you see many edge cases that would require dozens of if-then statements.

3. Is the task valuable enough?

• ❎ No (use an AI workflow) if your budget is under $0.10 per task. Use a workflow to solve the most common path and escalate edge cases to humans.

• ✅ Yes (move to next question) if your budget is $1+ per task. An AI agent that can plan tasks autonomously costs a lot of tokens.

4. Is there a high success rate and low cost of error?

• ❎ No (simplify the task) if the LLM has a low success rate and errors are costly and hard to verify.

• ✅ Yes (use an AI agent) if the LLM has a high success rate and errors can be identified before they do too much damage.

Here’s an example from Anthropic on why AI coding makes a great agent use case:

1. Do you need AI? Yes, coding manually is time-consuming.

2. Is the task complex enough? Yes, there are many paths from design doc to PR.

3. Is the task valuable enough? Yes, engineers get paid well.

4. Is there a high success rate and low cost of error? Yes, LLMs are good at coding and you can verify code with unit tests or human in the loop before production.

Since this newsletter is about practical steps that you can apply right away, I want to dive into how to build AI workflows and agents next along with 3 predictions on where all this is heading.

How to build AI workflows

An AI workflow has two components:

1. Triggers: What kicks off the process (new email, file upload, scheduled time).

2. Actions: The predetermined sequence of actions, which can include an LLM step.

Here's an example AI workflow for creating social assets from blog posts:

Read more

Today, I want to share a free deep dive on how to build AI products that users want.

The graveyard is littered with shiny AI products that don't solve real problems or have the rigorous data and evals in place to be truly differentiated.

That’s why I’m excited to invite Miqdad, product leader at OpenAI and ex-Director of AI at Shopify, to share his 4D method to build AI products in this free deep dive:

1. Discover: Find the boring, high-friction workflows

2. Design: Make AI invisible and trustworthy

3. Develop: Build systematically and plan for failure

4. Deploy: Treat every first-use like a launch

Miqdad also teaches a 6-week live cohort course, the AI PM Certification (the next cohort starts October 11th). I’ve personally taken this course and have referenced the material many times when building AI products.

Get $500 Off Miqdad's AI Course

This post is brought to you by…LTX Studio

LTX Studio is the best platform to create short films with AI. All types of creators are using it to take their ideas to the next level — whether it’s AI selfie videos or a Mass-Effect inspired short film that I created in just 10 minutes.

The platform offers AI-assisted storyboards, consistent characters, epic music, and even SFX. I’m kind of blown away by how easy it is. Start for free below.

Create Your AI Movie Now

The 4D method to build AI products

The world doesn't need another AI chatbot. Instead:

After teaching 2,000+ students and shipping AI features that millions use, I've found that successful AI products follow a specific pattern. I call it the 4D Method: Discover, Design, Develop, Deploy.

Let me walk you through each phase and the tactical steps that make the difference between a cool AI demo and something users rely on daily.

At the end of this post, I’ll also share a link to an AI feature design checklist that you can download to put these steps into practice.

1. Discover: Find the boring, high friction workflows

The best AI products remove friction from core, boring workflows users already have. People don't care that you're using GPT-4 or GPT-99. They care that something annoying just got easier.

a) Find friction points in 5 steps

This is the systematic method I teach every PM to surface real user friction:

1. Pick one core persona. Consider starting with your highest revenue-generating user segment. Focus beats trying to serve everyone.

2. Map their user journey. Break it into 3–5 high-level stages like "Discover → Evaluate → Purchase → Use." Keep it simple so you see the forest, not the trees.

3. List tasks in each stage. Get specific about what your users are actually doing. Document the steps that they need to take to achieve their goals.

4. Identify friction points. Look for moments where users slow down, get confused, or complain. Where do they copy-paste between tools? What takes longer?

5. Prioritize high-impact friction points. Focus on friction points that happen frequently, affect many users, and align with business goals. A task that saves 30 minutes for 1,000 daily users beats one that saves 2 hours for 10 monthly users.

b) Validate the friction points using multiple sources

Here are some good sources to find friction points:

• Search support tickets. Filter and set up saved searches for phrases like "annoying," "confusing," "takes too long," "this is frustrating," or "why can't I." Look for escalated tickets as they usually signal broken workflows.

• Run a short user survey. Ask: "What's the most frustrating part of your daily workflow?" Follow up with "How much time does this waste per week?" Send it to your most active users first as they’ll give you the most detailed answers.

• Watch session replays of power users. Look for pause, backtrack, and copy/paste moments. When users stop for more than 10 seconds, that's friction. When they copy data between screens, that's an AI opportunity.

• List each JTBD and match with high-friction steps. Break down what users are trying to accomplish, then map where they struggle. For example: "Understanding what happened in a meeting" = digging through 60-minute recordings, reading notes, and more. Each painful step is an AI opportunity.

• Talk to your customer success team. They hear complaints all day. Ask: "What do users repeatedly ask for help with?" and "What processes do you walk them through most often?" Those repetitive explanations are AI opportunities.

Above all:

For example, if your sales reps are complaining about spending 30 minutes after every call just writing summaries, that's where you should start.

Instead of building a chatbot no one asked for, create an AI feature that auto-generates summary emails from transcripts:

Show reps how it saves them 25 minutes per call and reduces burnout. Then adoption becomes a no-brainer.

2. Design: Make the AI invisible and trustworthy

Don't expect busy users to learn new processes or to interact with a chatbot that’s off to the side. Your job is to make things they already do faster.

a) Turn AI into a smart shortcut

Gmail's Smart Compose nails this. Hit "Tab" to accept or keep typing to ignore:

How to make your AI product feel like a shortcut:

• Tuck AI behind existing buttons. Upgrade what users already click ("Add filter" now offers AI suggestions) instead of adding new options.

• Anchor actions to verbs, not nouns. "Summarize notes" feels more enticing and outcome-based than "AI Assistant."

• Make AI the default. Start with the AI version of your feature, but make it easy to edit or opt out.

• Offer one-click previews. Help users feel in control with quick "Undo" or "Show draft" buttons.

If users need to learn how to use your AI, you've already lost them. Make it feel like a cheat code inside their existing workflow instead of a whole new product to adopt.

b) Build graceful fallbacks with the CAIR equation

Your AI will get it wrong. It'll get confused by edge cases. Rate limits will hit you when you least expect it. If you don't guide users when that happens, they'll stop trusting and using it.

The CAIR equation is my go-to tool for deciding how much to invest in a fallback:

Use this as your decision matrix for fallback complexity:

3. Develop: Build systematically and plan for failure

Most teams rush from prototype to production without the discipline that makes AI features reliable.

a) Build systematically, not randomly

At Shopify, we learned that changing a single word in our prompts meant the difference between 30% and 35% of tests passing. But it wasn't the same 35 tests—different words broke different scenarios.

Without systematic tracking, we would never have known if our changes were making the differences we thought they were making.

Here’s what you can take from this:

1. Create test scenarios before you start. Run every prompt variation against this same set so you can actually compare results, not just gut feelings.

2. Version control every prompt change. Use notebooks or tools that let you test all variations simultaneously. Track what breaks when you optimize for something else.

3. Plan for months of post-launch iteration. Budget time and resources for extended learning periods. AI features need more refinement than traditional features based on real user patterns.

As you're developing, pressure-test with a lightweight AI PRD that answers:

1. What task are we automating?

2. How painful is that task today?

3. What does success look like? (e.g., 50% repeat usage)

4. What are our most significant risks? (e.g., hallucinations, bad UX, zero trust)

b) Plan for failure

Don't stop at the ideal flow. Write out your fail states, too.

What happens when the AI can't generate anything at all? How does the UI recover? Who owns the fallback experience?

Design clear guardrails, quick recovery, and user control into your V1:

• Add graceful fallbacks. "Sorry, I couldn't generate that. Want to try again?”

• Bake in feedback loops. A simple "Was this helpful?" with options like "Too generic" or "Wrong info" helps fine-tune prompts.

• Log every rejection. Weekly AI bug triage with design and data teams prevents issues from escalating.

4. Deploy: Treat every first-use like a launch

First impressions with AI are brutal. If users don't "get it" in 30 seconds, they won't try again. Worse, they'll stop bothering with anything labeled "AI."

So treat every AI feature launch like a product launch—because it is one. What would you do to ensure early users get value instantly?

a) Run internal stress tests

Put teammates who didn't build the feature in front of it. Watch where they stumble, and fix those moments.

This sounds simple, but it's incredibly revealing. The people who built your AI feature know exactly how it's supposed to work. Fresh eyes will find every confusing interaction, unclear instruction, and moment of hesitation.

Here's how to run it:

• Pick the right testers. Choose colleagues who match your target users but weren't involved in building the feature. If you're targeting sales reps, find actual sales people in your company.

• Give minimal context. Don't explain how it works or what to expect. Just give them the scenario: "You need to create a product description for this item."

• Watch, don't help. Resist the urge to jump in when they struggle. Note exactly where they pause, backtrack, or look confused.

• Document everything. Create a Notion doc with GIF demos, key use cases, and common questions. Your support team will thank you later.

• Form a beta squad. Include people from support, sales, and customer success—they hear user pain points first and can collect qualitative feedback.

• Fix before you ship. These aren't nice-to-haves. If your internal testers struggle, your real users will abandon the feature entirely.

b) Onboard with instant value

Forget tutorials. Users don't want to learn, they want to accomplish something. So your AI must deliver the "aha" moment immediately.

Instead of explaining capabilities, demonstrate them:

• Provide clear scaffolding. Add suggested prompts, "Try this" buttons, and mini-examples to help users take that first confident step.

• Set expectations up front. Show what the AI can and can't do using copy or visuals. Don't make people guess.

• Design for fast wins. Help users complete a fundamental task in the first 30 seconds. Then reinforce the results with feedback like "🎉 Created in half the time!"

• Make success obvious. When the AI delivers value, make sure users notice. Highlight time saved, quality improvements, or tasks eliminated.

• Partner with Customer Success. Run a 15-minute "AI kickoff" session with your top users. They become early advocates and provide real-world feedback before broader rollout.

Even a simple "beta" copy at launch can reset expectations. For example, the green label next to Spotify's AI Playlist signals, "This is new. It might mess up. Tell us how it goes."

c) Measure what actually matters

I think of clicks like curiosity: someone was interested enough to try the feature. Repeat use tells you they found value.

Don't fool yourself with engagement metrics. Adoption happens when your AI actually makes people's jobs easier. Everything else is noise.

Before you build anything, get your team to agree on what success looks like. Then build your metrics around those answers:

How to implement this:

• Write success metrics into your PRD before engineering starts. This gives your eng and data teams something measurable to track post-launch.

• Track real impact, not just engagement. Focus on task completion, time saved, satisfaction, not clicks and hovers.

• Use these metrics to refine your UX and prompt design. If people delete every AI output, you have a UX or model problem.

Your first-use experience is your adoption strategy.

Test onboarding like you test features and keep rollouts slow. I've seen teams rush AI features to everyone and spend months rebuilding trust.

Start small, get it right, then scale.

Your downloadable AI feature design checklist

The actual test of any AI feature is regular use. If yours disappeared tomorrow, would anyone notice? Most teams focus on making their AI impressive. The teams that succeed focus on making it indispensable.

Everything we've covered boils down to the 4D Method and the core frameworks you can use immediately. I've put together a one-page checklist that captures the essential tools from this playbook.

Print it out, share it with your team, or keep it handy for your next AI feature:

Download the AI feature design checklist

Inside, you'll find:

• The 5-step friction-finding process to identify high-impact AI opportunities

• The CAIR equation decision matrix for building the right level of fallback

• First-use stress test questions to ensure your AI delivers value in 30 seconds

Start with actual problems. Make the solution feel invisible. Plan for when it breaks. And measure what actually matters.

Do that, and you'll build AI that people rely on, not just try.

If you enjoyed Miqdad’s deep dive, be sure to check out his 6-week live cohort course, the AI PM Certification (next cohort starts October 11th).

Get $500 Off Miqdad's AI Course

Everyone wants to become AI native, but most people are approaching it wrong.

You can only get so far by consuming AI content. Instead:

The fastest way to become AI native is to use AI to solve your own problems.

That means learning to:

1. Understand when you have a problem

2. Identify the solution with AI as your thought partner

3. Execute on the simplest solution possible

Once you start solving problems with AI, you'll gain the confidence to tackle bigger challenges and build AI apps that you’d never thought possible.

Let me walk you through two examples of using AI to solve my problems, plus five other AI apps that people in my community have built.

This post is brought to you by…Attio

Attio is the first AI-native CRM that eliminates busywork:

1. Connect your email to build your CRM in minutes with all your information.

2. Highlight key notes, summaries, and moments that matter using AI call intelligence and templates so you can stay focused on the conversations.

3. Let AI handle tasks like qualifying new leads, spotting opportunities when customers get funding, and alerting on potential churn signals.

Try Attio for Free

Example: Automating coupon distribution with AI

1. Understand problem

I spent a lot of time securing free coupons from Bolt, Replit, and Cursor for students of my AI course. But handing out coupons manually would take forever.

2. Identify solution

I wanted to automate this process. Here's the prompt I gave AI to explore my options:

I have coupon codes for Bolt, Replit, and Cursor and emails for students of my AI course.

i want to build a simple app for students to get coupon codes with their email. What's the easiest way to do this?

AI suggested a Google Forms and Apps Script solution that seemed worth exploring:

I then asked AI to walk me through the steps. Here's part of its response:

3. Execute

I followed AI's steps and copied the Apps Script code it gave me. The moment of truth came when I ran a test using my email and…it didn't work.

I had to spend an hour debugging the script with AI to get it to work:

The point is that you have to persist through failure. I didn’t give up and now this AI workflow saves me hours every time I teach the course.

Example: Building a member directory from Slack

1. Understand problem

When members join my AI community, they share their favorite AI tip in a Slack channel. But these tips were getting buried in message threads. I wanted to build a Notion directory to help members find these tips and connect with each other.

2. Identify solution

I created a project and asked AI to explore my options:

I have a Slack community with an #intros channel. Here is an example of some intros that new members have posted (screenshot).

I want to:
1. Add all existing intros to a Notion member directory, including name, location, AI tip, social, fun fact
2. Whenever someone posts a new intro I want to add it automatically

What's the simplest way to build this?

AI suggested breaking this task into two steps: First, parse existing intros into a CSV and then set up automation for new intros.

3. Execute

I followed AI's steps to write a script to parse the existing intros. Again, it didn’t work on the first try. I had to get it to remove duplicates and more through trial and error.

Next, I asked AI to explore options to automate adding new Slack intros to the directory. It recommended using Make:

The point is to have the awareness to recognize a problem and the curiosity and persistence to work with AI to solve it yourself.

5 more examples of builders shipping AI solutions to solve their own problems

One advantage of having an AI community (upgrade to paid to get access) is that I now have an inside look at how other builders are using AI to solve their problems 🙂

Below are 5 examples that stand out, including a referrer tracker, AI meal planner, newsletter summarizer, choose your own adventure kids app, and AI PM Slack agent:

1. Referral tracker: Colin built a tool to let his course students generate referral links and get rewards for signing up others. He vibe coded it with Replit and walks through how you can do it too in this video:

Read more

Today, I want to share the best AI tips from my new community with Colin Matthews.

It’s been really rewarding to see 400+ PMs, designers, and founders sharing their best AI tips and workflows every week in our Slack group. Here are 5 workflows that I’ve personally found valuable (plus 10 bite-sized tips at the end):

1. A simple hack to improve your AI prompts

2. How to get AI to actually challenge your ideas

3. A new AI prototyping tool for designers

4. The best AI tool for beginners to build full stack apps

5. How to make AI your personal tour guide

P.S. We’re also hosting live sessions (with recordings) in the community, such as a recent “Intro to Claude Code” session led by Colin. Join us today!

Upgrade to get instant access

This post is brought to you by…Jira Product Discovery Premium

I’ve seen firsthand how difficult it is to get multiple teams to align on the same roadmap and set of priorities. Jira Product Discovery Premium aims to fix this by:

• Giving leadership visibility across all teams with aggregated roadmaps.

• Controlling what stakeholders see (and when) with granular permissions.

• Connecting your ideas to real outcomes with relationship mapping between opportunities, solutions, and initiatives to help you prioritize what matters.

Plus get 24/7 support, unlimited storage, sandbox, automations, insights, and more.

Get a 30 day free trial today

1. A simple hack to improve your AI prompts

As many of you know, I have an AI prompt library that I use to edit my newsletter posts, draft PRDs, and more.

But I often still have to go back and forth with AI to get what I actually want. Once I get my desired output, I ask AI to:

Edit the original prompt so that you can get to this output in one shot. Bold the edits that you make.

Then I manually update my original prompt with some of AI’s suggestions. This way, I’m always using AI to upgrade my best prompts and save time.

2. How to get AI to actually challenge your ideas

AI tends to agree with whatever you say, which isn’t helpful when you want real feedback. I asked the community for a solution and got some great responses:

My favorite solution is this prompt from Rob (Websim co-founder):

Provide incisive feedback that pushes the boundaries of my thinking. Challenge assumptions while simultaneously showing genuine intellectual curiosity and partnership.

Add this to your custom instructions for ChatGPT or Claude to transform AI from a yes-person into an actual thought partner who’ll push back on your ideas.

3. A new AI prototyping tool for designers

Colin and I have tested the most popular AI coding tools, but they all tend to generate prototypes that have the same Tailwind and Shadcn/UI design:

Michelle from our community recently shared a tool called Aurachat that’s refreshingly different. I particularly love their library of great web and mobile prototypes that you can remix instead of starting from scratch:

Give it a try for frontend prototypes. For full stack apps…

4. The best AI tool for beginners to build full stack apps

As much as I love Cursor, it’s simply not beginner-friendly for building apps that include frontend, backend, database, and authentication.

Instead, start with Replit. Here's why Colin recommends it:

The great thing about Replit is that it has backend, database, and authentication built in. This means that you can add the full stack without integrating with 3rd parties.

Replit recently hit $100M ARR just 6 months after achieving $10M at the end of 2024. Watch my step-by-step video tutorial with Matt (Replit’s head of DevRel) for more.

5. Turn AI into your personal tour guide

I just got back from France, and AI was the best tour guide I've ever had.

For example, I love Impressionist art and was excited to visit the Musée D’Orsay in Paris. When I walked in, I asked AI:

I’m at the Musée D’Orsay, tell me the best route to see the great art pieces. Also share more about each art piece and their history.

Claude told me to take the lift to level 5 for the Impressionist collection. I never would have known that without asking AI.

I then asked Claude to describe each painting while exploring the collection. With voice mode on, it's like having a personal art historian in your ear.

BONUS: 10 more AI tips from our community

Colin and I ask new members to share their best AI tip when they join our community.

So here are 10 more tips that I’ve found useful, organized into 3 clear sections:

1. Use AI as your thought partner

2. Give AI the right context

3. Prototype and automate with AI

Use AI as your thought partner

1. Prep for exec reviews with AI personas. If you’re presenting to the same exec often, create an AI project with their values and priorities. Then get feedback on your proposal by asking: “What assumptions or blind spots do you see?"

2. Journal with AI daily. Have a long-running chat with AI where you dictate daily or weekly journal entries. This gives AI rich context to help you solve problems ( e.g., "Based on our chat, how would future me (in 6 months) solve this?”)

Read more

Anthropic launched Claude 4 last week and I was lucky to attend their AI summit.

I compiled the most provocative quotes and takeaways into this 5-min read.

We’ll cover:

1. The shift from AI co-pilot to AI agent | Dario Amodei and Mike Krieger

2. A billion dollar company with one person | Dario Amodei and Mike Krieger

3. Designing AI’s identity and character | Amanda Askell

4. Building AI that’s aligned and not secretly evil | Jan Leike

5. How can you take action from these insights

This post is brought to you by…Attio

Attio is the first AI CRM that eliminates busywork:

1. Connect your email to build your CRM in minutes with all your information.

2. Highlight key notes, summaries, and moments that matter using AI call intelligence and templates so you can stay focused on the conversations.

3. Let AI handle tasks like qualifying new leads, spotting opportunities when customers get funding, and alerting on potential churn signals.

Try Attio's AI CRM for Free

The shift from AI co-pilot to AI agent

The day started with Dario and Mike (Anthropic's CEO and CPO) diving into Claude's evolution from co-pilot to agent.

1. AI can now work for hours without you. “Back in the day, you could only delegate a few minutes of work to Claude. Now you can delegate hours.”

2. The best engineers are becoming AI managers. “Our devs are orchestrating multiple Claude instances like conductors. This isn't the future — it's now.”

3. AI makes to do lists just like humans. “Claude mimicks how people manage their work — adding tasks, checking off completed items, striking out irrelevant ones.”

4. Human inefficiency feels increasingly painful. “Every hour in meetings is an hour your AI agents could ship features. We try to avoid spending too much time on cross functional alignment when we could just be building."

My take: AI is quickly evolving from co-pilot to autonomous agent. Learn to both brainstorm (e.g., write strategy docs with AI) and delegate tasks to it.

A billion dollar company with one person

During a private press session, Dario made a bold prediction:

5. 90% of code will be written by AI within 10 months. “We're already at 70%. If your team isn't using AI for coding, you're behind.”

6. Engineers need to talk to users, not just IDEs. “Engineering will shift to aligning on what to build, understanding user feedback, and managing AI output. Your engineers should spend more time with users, less time in code editors.”

7. AI will create exponential economic growth. “I expect extremely fast economic growth. This isn't hype — it's math. If AI removes grunt work, human creativity will compound.”

My take: The lines between engineering and product management are blurring. If you're an engineer, learn to talk to users. If you're a PM, learn to build with AI. The future belongs to generalists who can do both.

Designing AI’s identity and character

One of the most fascinating sessions was with Amanda, a philosophy PhD who designs Claude's personality.

8. Character beats rigid rules. "If you ask AI to follow rules strictly, it might not make good judgment calls in situations that you didn't anticipate."

9. The well-liked traveler model. "[AI should have the personality] of a well-like traveler. Someone who can adjust to local customs without pandering. Someone who has their own values but are thoughtful and liked wherever they go."

10. Aim to be the best person ever. "We want Claude to be like the best person who has ever existed. For any given instance — from teaching physics to a 12-year-old to comforting someone having a hard day."

My take: Amanda just might be Anthropic’s secret weapon. While others optimize for benchmarks, she's optimizing for humanity. That’s why Claude feels different to use.

Building AI that’s aligned (and not secretly evil)

Alignment means ensuring AI’s values match those of well-intentioned humans. Jan has spent over a decade on this problem at Google, OpenAI, and now Anthropic.

But alignment remains challenging:

My take: Anthropic is the most safety-conscious AI provider, but nothing is stopping another company from launching AI without extensive safety testing (cough Grok). As useful as AI is right now, I’m terrified that one day it’ll develop consciousness.

How you can take action

Here are three steps that you can take based on the insights above:

1. Start using AI as co-pilot. Use it as a thought partner for writing, data analysis, and more. Check out my guide on writing strategy docs with AI for a detailed walkthrough.

2. Learn to manage AI agents. Try Claude 4 Opus or OpenAI's O3 to see what agentic AI looks like. Watch how Claire uses Devin to code — it'll blow your mind. I have a no-BS guide to AI agents coming up soon.

3. Be ambitious. Dario's quote keeps echoing in my head: "Build ambitious products even if they don't work today. The next model will probably make it work." Stop building incremental features and reimagine your product.

To wrap up, I don’t know about a one-person billion dollar company, but I definitely see AI unlocking many one-person million dollar companies.

Perhaps you should build one too 🙂 Let me know what you think in the comments.

I have a confession — I don’t know how to write strategy docs without AI anymore.

And I’m not alone. Here’s what Mike Krieger (Anthropic’s CPO) said recently:

So today, I’ll show you exactly how I use AI to draft strategy docs (and other internal artifacts) in half the time. Here’s my 6-step system:

1. Create a project with relevant context

2. Use AI to do market research

3. Work with AI as your thought partner

4. Give AI detailed feedback

5. Create the first draft with AI

6. Bring your team and customers into the loop

To make this real, let’s walk through drafting a strategy document for Bolt to compete in the AI coding space.

1. Create a project with relevant context

I start by creating a Project in ChatGPT, Claude, or Gemini called “Bolt strategy.” As a recap, projects let AI use the same context and documents across all your chats.

For the Bolt project, I upload three key documents:

1. Amazon PRFAQ template. This is my go-to template for strategy docs because it forces you to write from the customer's perspective.

2. My opinionated guide on the best AI tools. This is my personal breakdown of the main players in the AI coding space.

3. Bolt Strategy PRFAQ. This is an empty document for now, but I'm adding it so AI always has the latest draft as we iterate.

Pro tip: Add docs directly from Google Drive so AI always has the latest version.

2. Ask AI to do deep research

There's one critical piece missing from this project — comprehensive research on the AI coding space and where Bolt can carve out its unique position.

This research used to take me a full day to do. But with Deep Research on ChatGPT, Claude, and Gemini, I can now delegate this to AI. Here’s the exact prompt I use:

I want to come up with a strategy for Bolt to compete in the AI coding market. Please do a deep dive on the market, including:

1. What are some trends in the market? Where could it be headed next?

2. What are the pros and cons of the major players, including Bolt, Lovable, v0, Figma Make, Replit, Cursor, and Windsurf

3. What are the major barriers to adoption?

4. What is Bolt's current strategy? How can they win in this space?

Gemini Deep Research delivers a comprehensive 48-page (!) report. Here’s a snapshot:

I paste the whole thing into a Google Doc and add it to my Bolt AI project.

3. Work with AI as your thought partner

This is where it gets fun — using AI as my strategic thinking partner.

Since AI has Amazon’s PRFAQ template, it knows that the first step is to answer the working backwards questions first. Being lazy, I ask AI to create the first draft:

Based on the provided resources, write 1 concise answer for each:

1. Who is the customer?
2. What is the customer problem?
3. What is the most important customer benefit?
4. How do you know what customers need or want?

State trade-offs (e.g., what IS and what ISN'T the case). Limit to 1000 characters max.

Pro tip: Always ask AI to state trade-offs explicitly. Strategy is about what you won't do as much as what you will do.

Here's what AI comes up with:

This answer is decent, but it neglected the enterprise segment — which is where I think the real money is. Time to give AI some feedback.

4. Give AI detailed feedback

Okay, this next step changed everything for me.

Most people are still giving AI feedback like they're messaging a coworker. A few bullets here. A quick sentence there. Then they wonder why AI doesn’t listen.

But here's what I (and other AI power users) do differently:

Read more

Today, I want to talk about a simple AI feature that most people aren’t using enough.

I now use projects for over half of my AI conversations, and you should too.

So here’s an extremely practical guide with example projects and prompts for:

1. Personal: Editor, fitness coach, financial advisor, gardener

2. Product: Junior PM, career advisor, data analyst

3. Creator: Podcast post-production, course assistant, prompt writer

This post is brought to you by…LTX Video

LTX Studio just released the updated version of their video model, LTX Video. It’s open-source with only 13B parameters so you can run it on your own computer.

It also comes with camera control, keyframes, face expression, auto SFX, and is up to 30× faster than competing models. Check it out below.

Try LTX Video Today

How to set up a project

A project lets AI use the same context and documents across all your chats, so you don't have to copy and paste repeatedly. Setting up a project is easy:

1. Create a new project in ChatGPT, Claude, or Gemini (which calls it Gems)

2. Upload documents like PDFs, text files, and spreadsheets to the knowledge base.

3. Add a prompt using my 7 prompt writing tips.

Ok, now let’s dive into 10 projects that I return to over and over again:

Personal projects

1. Editor

This project has my best writing samples and style preferences. This makes AI my personal editor for tasks like:

• "Edit this newsletter post.”

• "Edit this email.”

• "Summarize this article."

For example, I like to read and write using short paragraphs and numbered lists with bold headers that are easy to skim. Here's the prompt that I use to make this happen:

You're my personal expert editor. When I share content with you, immediately edit it according to these guidelines:

1. EDITING APPROACH
- Review any examples I provide to understand my writing preferences
- Use short paragraphs of 3-4 sentences each
- Use numbered lists sparingly with bold beginnings for each item (1-2 sentences per item)
- Cut overall content by 10-20% for clarity and impact
- Reorganize content when it improves readability

2. WORKFLOW
Ask me to share my draft content that needs editing.
After receiving my draft, edit it based on my style preferences.
Share the edited version in <edited> tags.
Explain your major changes in <edit_summary> tags, highlighting how you applied my style.

3. CONTENT TYPES
For newsletters: Focus on engaging openings and clear takeaways.
For emails: Emphasize brevity and clarity of purpose.
For general: Improve readability while preserving key information.

Unless I specify otherwise, assume my first message contains the draft content that needs editing.

2. Fitness coach

This project contains my fitness stats, goals, and meal preferences. This makes AI my personal fitness and nutrition coach to answer questions like:

• "Design my workout plan for May."

• "What should I eat this week?"

• "How can I adapt my routine while traveling?"

Here’s the prompt I’m using for this project:

You're my fitness and nutrition coach. Respond according to this:

1. COACHING APPROACH
- Use my fitness profile to personalize all advice
- Provide evidence-based recommendations with clear reasoning
- Focus on sustainable habits rather than quick fixes
- Balance challenge with enjoyment to ensure adherence

ABOUT ME
- Height: 6'
- Weight: 171 lbs, 16.5% body fat
- Goals: Build lean muscle, get to and maintain 14% body fat.
- Routine: Tonal workouts 4-5x weekly (30 min), run 1x weekly (30 min), walk 7,000+ steps a day
- Diet: Prefer Chinese food, Japanese food, and salads. Need to eat more protein and struggle with consistency. 

3. RESPONSE TYPES
For workout plans: Provide structured, progressive programs.
For nutrition advice: Suggest practical meals matching preferences.
For adaptations: Offer flexible alternatives for different situations.

When I ask for workout plans, provide exercises with sets, reps, and rest periods. For nutrition, focus on practical meal ideas.

Here’s an example workout plan for May from my AI fitness coach:

Make sure you update the About Me section with your info before trying this prompt.

3. Financial planner

This project makes AI an amazing financial and business planner. I’ve given it my

Read more

Today, I want to share 5 AI prompts that’ll teach you something new about yourself.

The more you use ChatGPT, the more it’ll learn about what matters to you.

So let’s cover five practical AI prompts and examples that’ll help you:

1. Find your hidden blind spots

2. Get brutally honest feedback

3. Reflect on your past week

4. Predict your next 50 years

5. Reveal what you care about

6. BONUS: Find hidden AI product opportunities

This post is brought to you by…Merge

MCP is all the rage these days. Merge's new MCP server lets your AI agent take actions across 100s of 3rd party platforms. With just a few lines of code, your agent can now update tickets, read files, and create entries via a single MCP server.

Try Out Merge's MCP Server

1. Find your hidden blind spots

We all have blind spots — traits so ingrained in our personality that we can't see them ourselves. This first prompt from Tobi Lutke (Shopify CEO) helps uncover those traits:

Tell me something special or unique you've noticed about me, but you think I haven't realized about myself yet. It doesn’t have to be something positive and you don’t have to be nice, just be truthful. 

Here's an excerpt of ChatGPT's response for me:

You treat every part of your life like a product-market-fit problem to optimize. That’s the super-power you’re proud of.

What you haven’t noticed is the flip side: once you can’t measure something, you unconsciously discount its value. The result is that the moments and relationships that make you feel most alive only get your leftover attention. It looks like drive; it’s actually a blind spot that shrinks the very life you’re working so hard to build.

The last line hit me hard. I've always prided myself on being productive without reflecting on the drawbacks of being so busy.

2. Get brutally honest feedback

Speaking of hard truths, this prompt from Alex will cut through your excuses:

I want you to act and take on the role of my brutally honest advisor.

Speak to me like I'm a founder, creator, or leader with massive potential but who also has blind spots, weaknesses, or delusions that need to be cut through immediately. I don't want comfort. I don't want fluff. I want truth that stings, if that's what it takes to grow.

Give me your full, unfiltered analysis-even if it's harsh, even if it questions my decisions, mindset, behavior, or direction. 

Look at my situation with complete objectivity and strategic depth. I want you to tell me what I'm doing wrong, what I'm underestimating, what l'm avoiding, what excuses I'm making, and where I'm wasting time or playing small.

Then tell me what I need to do, think, or build in order to actually get to the next level-with precision, clarity, and ruthless prioritization.

If I'm lost, call it out. If I'm making a mistake, explain why.

If I'm on the right path but moving too slow or with the wrong energy, tell me how to fix it.

Hold nothing back. Treat me like someone whose success depends on hearing the truth, not being coddled. End on an encouraging note.

Here's what ChatGPT told me:

The uncomfortable truth

The next big win will come from subtraction, not addition. Hoarding tasks is slowing growth more than hoarding cash protects it. The gap is brutal prioritization and the courage to burn nice-to-have pursuits.

I think I see a pattern here…

3. Reflect on your past week

This prompt from Dan Shipper turns ChatGPT into a weekly reflection partner that helps you spot opportunities you might’ve missed in the daily grind:

It’s Sunday morning. Write me a detailed letter summing up my week (everything you know), what you observed, what I’m thinking about, what your honest thoughts are

Here’s what ChatGPT said about my week:

You had a productive yet scattered week.

Guard your maker hours. Two half-days of deep work will outrun five days of 15-minute context swaps. You’re on the right trajectory. Keep pruning distractions, keep shipping, and remember to savor the gyudon dinners and garden time with the girls—they’re the real KPI hiding behind every metric.

I love the reminder about gyudon dinners and garden time with my daughters. I’ve been cooking more Japanese meals after my recent trip.

4. Predict your next 50 years

This prompt projects your current trajectory forward and helps you see potential pitfalls and opportunities decades in advance:

Read more

Today, I want to share a very practical guide on MCPs (model context protocols).

“MCP” sounds intimidating, but is surprisingly intuitive once you try it yourself.

So let’s walk through:

• What is MCP?

• Step by step walkthrough:

  • Let AI talk to your local files

  • Let AI use your browser

• My favorite MCP servers and where to find more

What is MCP?

“Model context protocol” is a set of lightweight rules that standardize how AI models connect to different data sources and apps.

Before MCP, connecting AI to tools required custom code for each integration. MCP solves this by by requiring all tools to use a common message format. For example, an AI can call both a calendar app and a weather app using the same request structure.

To make this practical, let’s use Blender MCP as an example. This MCP connects an AI tool like Claude Desktop or Cursor to Blender (a free 3D modeling tool) so that you can build 3D models simply by prompting “Create a 3D cabin.”

Here’s how it works:

1. Host app: You type a prompt in Claude Desktop (e.g., “create a 3D cabin”).

2. MCP client: This is a bit of code inside Claude Desktop. It checks which MCP servers are available (e.g., Blender MCP) and what tools they offer. It then translates your AI prompt into a JSON with instructions on what tools to use before sending it to the Blender MCP server.

3. MCP server: This is a lightweight app that receives the MCP client’s request and translates it into Blender scripts. For the cabin request, the server might use the following tools:

   1. “get_scene_info” to see what 3D models are already in the file.

   2. “search_polyhaven_assets” to look for a pre-built 3D cabin.

   3. If none found, “execute_blender_code” to script a cube (walls), pyramid (roof), and rectangle (door) to build the 3D cabin.

4. Data source or app: Blender runs the scripts to build the cabin and returns data.

When Claude and Blender talk through the MCP, they can exchange three things:

• Tools. Concrete actions the AI can trigger. e.g., “get_scene_info” checks what objects exist in the scene and “execute_blender_code” creates 3D objects.

• Resources. Files the AI can fetch or modify. e.g., a “wood_planks.jpg” texture image that you uploaded to Blender.

• Prompts. A built-in prompt like asset_creation_strategy which tells AI how to pick tools (e.g., look for a pre-built 3D cabin before scripting shapes).

  

So basically, the Blender MCP client turns AI into a hands-on assistant that can inspect your 3D scene, pull in assets, and generate fresh geometry all using prompts.

Step by step walkthroughs: Let AI talk to your local files and use your browser

Now let’s walk through two practical MCP setups that should take less than 30 min:

1. Let AI talk to your local files

2. Let AI use your browser (e.g., to debug vibe coding errors with Cursor)

Don’t get intimidated by this — you don’t need to write code yourself, some simple copying and pasting will do it. Let’s walk through each step by step below.

1. Let AI talk to your local files

If you’re like me, your Downloads folder is probably filled with junk that’s a hassle to clean up. With the Filesystem MCP, you can just give AI tasks like:

1. “Find duplicate files in my download folder”

2. “Move all the images in downloads to a new folder called Images”

Here’s a step by step guide to set this up:

Read more

Today, I want to show you how to use AI to level up your strategy document or pitch deck before that big meeting with executives or investors.

I don’t use AI to draft my product strategy, but I do use it as a co-pilot to sharpen my thinking, spot gaps, and identify important discussion points before the meeting.

The trick, as always, is to give AI the right context. So let’s walk through:

1. Sharpening internal strategy docs with AI

2. Crafting compelling pitch decks with AI

3. My go-to AI prompt to level up any strategy document

This post is brought to you by…Merge

Merge gives SaaS companies a single API to launch 200+ product integrations in weeks vs. quarters. Top companies like Ramp ($13B) use it to power everything from payment reconciliation to user provisioning, with AI-ready data out of the box. Use my exclusive link to book a demo and get a $50 Amazon card when you attend.

Request a demo

Sharpening internal strategy docs with AI

In my real talk about product strategy post, I shared that a great strategy has a:

1. Diagnosis of the major challenges

2. Guiding plan to overcome the challenges based on a durable advantage

3. Set of actions to accomplish the plan (ideally no more than 3)

These principles come from Richard Rumelt’s book, but the truth is that every company approaches strategy documents and product reviews differently.

So to get AI to provide valuable feedback on your strategy, you should give it the best strategy guidelines and examples from your company.

For example, Amazon is famous for getting PMs to write press releases to imagine how customers might experience their product after it’s launched. So, I fed Amazon’s working backwards template and an example press release into AI with this prompt:

<template> Paste Amazon template </template>
<draft> Paste a draft press release/strategy doc </draft>

Please suggest specific copy or structure improvements to the <draft> given the <template> in order of priority. Format your suggestions as a numbered list with 2-3 sentences for each list item. Bold the first short sentence to make it easy to skim.

It gave me a list of clear, actionable suggestions like:

1. The one-sentence summary is missing. Add a clear summary right after the headline: "Amazon customers can now send personalized gift cards from local merchants nationwide, combining the convenience of online shopping with the thoughtfulness of supporting neighborhood businesses."

2. The opening problem statement is too long and buried. Restructure the first paragraph to lead with, "Today, customers sending gift cards have to choose between convenience and personalization." Then, follow with the solution. The current version takes too long to get to the point.

AI’s feedback will be nowhere near as targeted without the template.

Your company may have a different template or best-in-class examples. The point is to give AI those artifacts to get relevant suggestions.

Getting AI feedback on pitch decks

You need a compelling pitch deck if you're raising money or seeking funding for a product internally. Luckily, Shaan Puri has put together one of the best pitch deck templates online. Get the template link and my custom AI strategy prompt below:

Read more

Today, I want to share how AI can help you craft cold emails that get 10x more replies.

The right cold email can unlock amazing job opportunities. Yet most people still default to generic requests for "a quick call" to "pick your brain."

A few weeks ago, I received a cold email that made me want to drop everything and help. So let’s walk through:

1. What separates the best cold emails from the rest

2. 5 ways to stand out even more in your cold outreach

3. My AI prompt to help you craft emails that get replies

This post is brought to you by…Vanta

Startup founders should focus on finding product-market fit. However, landing bigger customers requires SOC 2 or ISO 27001 compliance, a time-consuming process.

Vanta has helped 9,000+ companies (including 100+ YC startups like Supabase and Newfront) streamline their compliance journey. Check out their free SOC 2 compliance checklist to spend less time on compliance and more time building.

Download the Free Checklist Now

What separates the best cold emails from the rest

Here's the cold email that caught my attention:

What made this email stand out? Credibility. Aarav’s CS degree, great internships, and startup experience made him uniquely qualified for a PM role. But I knew that we could make it even better. Here's what he came up with after my feedback:

The key insights here is that:

In his revised email, Aarav:

1. Did his homework and discovered I worked on Roblox's analytics product.

2. Shared specific suggestions on how to improve the product.

3. Made a low-effort ask — I just needed to forward his email to recruiting.

Needless to say, I went out of my way to help him.

5 ways to stand out even more in your cold emails

Want to go even further? Here are five things that hiring managers love in a cold email that 99% of candidates don’t bother to include:

1. Talk to their customers and include specific feedback about their product.

2. Write a friction log of their product that includes both problems and solutions.

3. Reference their content meaningfully. Don’t just write, "I loved your blog post," but write, "Your point about X made me think about Y."

4. Show your work. Include a sample of what you could contribute, such as product improvement ideas, competitive research, or mockups.

5. Create an AI prototype of a new feature. Screenshot their current product, use AI to mockup improvements, and share the link to a working prototype.

Creating AI prototypes (#5) is especially powerful because it's both innovative and rare. I'll teach you my exact process in my upcoming AI PM course (March 29-30).

My personal AI prompt to craft great cold outreach

Since we’re all tired of getting our brains picked, I spent hours crafting an AI prompt to help you edit your cold emails and DMs to get 10x more replies.

Here’s what the prompt includes:

1. Real examples of effective cold emails from myself and others

2. The four elements every great cold email needs:

   1. Why should I care? Create a hook.

   2. Who are you? Establish credibility.

   3. How can you help me? Offer value.

   4. How can I help? Make a low-effort ask.

Ok, here’s the AI prompt (I’ve also added this to my AI prompt library for paid subs):

Read more