Today, I want to show you how to build Claude Skills that catch their own mistakes and get better the more you use them.

I’ve become completely AI-skill pilled. I love encoding my knowledge and taste into skills and then reusing them to save time every week.

Watch my video now to see me build a skill from scratch in 5 steps complete with an eval loop and memory.

Timestamps:

• (00:00) What an AI skill is and the 5 steps

• (01:56) Step 1: Give AI your personal context and examples

• (05:00) Step 2: Edit the description to trigger the skill reliably

• (05:59) Step 3: Build an eval loop to have AI fix its own mistakes

• (11:54) Step 4: Add memory so the skill improves itself over time

• (13:40) Step 5: Build a skill editor to improve all your skills

• (17:44) Where human taste still plays a role

I’m proud to partner with Linear

As teams ship faster with AI agents, the bottleneck shifts from building to deciding what to build.

Linear’s new Agent understands your full workspace context: roadmap, issues, customer requests, and code. Ask it to surface patterns across feedback, scope out a spec, or catch you up on team progress.

Try Linear for Free

A quick recap of AI skills

As I shared before, a skill is just a folder with instructions that AI can trigger for a given task. In this tutorial, I’m going to build an /edit-post skill that can edit any newsletter or long-form piece. Here’s what the skill folder will look like at the end:

edit-post/
├ skill.md
├ example-tutorial.md
├ example-personal.md
├ example-product.md
├ evals.md
├ memory.md

1. skill.md tells AI what the skill does and when to trigger it.

2. The example files include my best posts so the skill can learn my writing style.

3. evals.md will have pass/fail checks the skill uses to grade its own output.

4. memory.md will log lessons from past chats so the skill gets better over time.

Now let’s build these files in 5 steps.

1. Create the skill using your examples and personal context

My newsletter has three types of posts: tutorials like this one, personal essays on career and life, and deep dives on AI trends and concepts. I saved my best posts for each type in separate md (text) files below:

Then I gave Claude Code all three example files and this prompt:

The most important thing here is to give AI as much personal context and examples of good output as you can. You should also keep these context files separate from the main skill.md for two reasons:

1. It keeps the skill.md lean because AI can decide which example posts to load (e.g., tutorial vs. personal post) based on your draft.

2. It lets you share the skill more easily without revealing your personal posts or information.

2. Be explicit about when AI should trigger the skill

When you ask AI a question, it doesn’t load the whole skill by default. Instead, it only reads the name and description to decide whether to load the full skill:

So it’s very important to include explicit “Use when…” instructions so AI knows when to trigger the skill automatically. You can always trigger it by hand too (type /edit-post).

3. Test the skill manually, then build an evals.md loop

Once you have a skill drafted, test it manually a few times.

For example, I ran /edit-post on a real draft, then gave it feedback to improve. After a few manual iterations, tell AI:

An eval is just written instructions that asks the AI to check its own work. I prefer pass/fail over scoring evals (like 4/5) because AI can’t reliably tell a 3/5 from a 4/5. Here’s a snippet of the evals that it generated for me:

Before we run the evals, give AI one more prompt:

This sets up the evals loop where one agent edits the post, a second agent grades it, and because the grader has a clean context window, it isn’t biased by the first agent’s work. If anything fails, the two will keep working until everything passes:

When I ran the loop on a newsletter draft, it took five rounds between the two agents to pass all the evals. AI stripped every em dash and “X, not Y” phrase along the way.

I think setting up this evals loop makes a big difference in leveling up your skills. You can just kick it off, go get coffee, and come back to a cleaner draft. You should still review and edit the final output manually of course.

4. Give the skill a memory.md so it improves over time

Evals improve the skill’s output and memory improves the skill itself.

To set it up, ask AI:

A memory.md is just a text file with a reverse-chronological log of what the skill learned from your past chats. Here’s an example:

Setting this up is optional, but I’ve found it useful for feedback that doesn’t fit a clean pass/fail check, like “make the voice more authentic.” Make sure skill.md references both evals.md and memory.md.

5. Build a skill that builds skills

I’ve been skillmaxxing for days, and I do worry that relying on AI to write skills could lead to a mess of slop that drags down the model’s response quality.

That’s why I built /skill-editor and /no-ai-slop skills to make all my skills concise and strip AI slop like em dashes, “X, not Y” phrasing, and duplicate instructions. Here’s a snapshot of my /no-ai-slop skill:

When I ran these skills on the /edit-post skill we just built, it caught a few em dashes, removed some redundant formatting, and made the skill more concise.

Quick plug: I’ve made this full post free, but if you want to copy my /skill-editor and /no-ai-slop skills, along with a dozen other handcrafted skills, consider becoming a paid subscriber to unlock all my skills and prompts at behindthecraft.com.

Follow these 5 steps to build skills that improve themselves

Here’s how you put this guide into practice. First, reflect on your past week and brainstorm with AI workflows that you can streamline with skills. Then create each skill by following these 5 steps:

1. Give AI your personal context and best-in-class examples to draft the skill.

2. Trigger the skill reliably by including “use when XYZ” in the description.

3. Build an evals.md with pass/fail checks so AI can loop to improve its own output.

4. Add a memory.md with lessons from your chats so the skill improves over time.

5. Build (or copy) a /skill-editor skill to keep all your skills clean and concise.

Watch the video now and subscribe to my channel if you enjoyed this tutorial. Let me know if you have any questions in the comments!

Today, I want to share a new episode with Josh Pigford.

Josh sold his last startup for $4M and is now building 5 products solo with agents. He showed me his exact AI skills stack, including a 4-step /build skill, an /adversarial-code-review skill where he pits Opus against GPT, a /but-for-real skill that makes the AI catch its own mistakes, and a /learnings skill that improves AI’s output over time. Josh has been a solo builder for 25 years and it was fascinating to watch how he works.

Watch now on YouTube, Apple, and Spotify.

Josh and I talked about:

• (00:00) Why launching fast beats months of planning

• (01:02) A tour of the 5 AI products Josh built solo

• (05:22) /build skill: Research, plan, track, implement

• (10:52) /adversarial-code-review skill: Getting GPT to review Opus' code

• (16:02) /learnings skill: Make AI update its own CLAUDE.md

• (16:47) /but-for-real skill: Make AI fix its own mistakes

• (19:11) Live demo: How to design non AI slop websites

• (28:55) Advice for builders without technical experience

If you enjoyed my chat with Josh, check out his new community for AI builders.

I’m proud to partner with Wispr Flow

There are many AI voice dictation tools out there but I’ve found Wispr Flow to be the most accurate at transcribing words. My kids love using it to dictate to Codex and Claude Code to build games and apps. Try it out with my free link below.

Try Wispr Flow for Free

Top 10 takeaways I learned from this episode

1. Keep shipping even though the fear of being embarrassed never goes away. After 25 years as a solo builder, Josh still thinks “It’s terrifying launching something. Every single time it’s just like, what if zero people care?” But he ships anyway: “The idea of spending months working on something before you put it out for other people to use, I think that’s a real bad idea.”

2. Use Josh’s /but-for-real skill to force AI to check for mistakes thoroughly before pushing a change. Josh finds that the skill surfaces at least 3-5 additional bugs that AI missed each time he runs it. The skill (see snippet above) is hilarious because it bullies the AI to work harder. You can find the full skill to copy at the end of this post and in my new skills library at behindthecraft.com.

Read more

Today, I want to show you how to build beautiful slide decks with Claude Code.

Instead of making PowerPoint and Google Slides manually, I built a /slides skill that turns a rough outline into a fully animated HTML deck in minutes.

Just install the skill and type “/slides (your rough outline)” to generate a beautiful slide deck. Watch my tutorial now for the complete walkthrough.

Timestamps:

• (0:00) Why building slides with AI is better than PowerPoint

• (1:30) My /slides skill: 12 formats and 3 templates

• (4:20) The 6 steps that make the skill actually work

• (5:16) Watch me generate a full deck in 3 minutes

• (8:13) What's actually inside my /slides skill file

• (11:57) The one lesson that matters when building with AI

I’m proud to partner with Oceans Talent

I hired Char through Oceans a few months back to help with podcast production. What I didn’t expect was how quickly he’d take it further. He’s now creating clips and picking up Claude Code and Codex on his own to deliver more every week.

Oceans Talent places operators, not assistants. Their talent is AI-fluent and delivers what a senior US hire would at 3-5x less cost. If you need help with marketing, ops, finance, or an EA, book a call at oceanstalent.com/peter.

Get Started for Free

Why you should build slides in HTML

Getting AI to generate slides in HTML has many advantages, including being able to:

1. Go from 0 to draft deck in minutes with a rough outline.

2. Create live, interactive charts with Chart.js.

3. Drop in images and have AI automatically resize them.

4. Add subtle motion to capture attention.

5. Run a visual QA loop where the AI polishes the slide for you.

6. Ship with zero dependencies since all slides render in a single HTML file.

12 beautiful slide formats and 3 templates

My /slides skill comes with 12 ready-to-go formats (cover, agenda, two-column, stat grid, feature bento, comparison, timeline, chart, and more) and three visual templates:

1. Default: Editorial and warm with a sharp red accent.

2. Dark: Bold and contemporary with a black background and yellow accents.

3. Light: Apple-style clean and minimal.

Here’s an example chart slide generated using the dark template:

How to start using the /slides skill now

You have two options:

1. Copy and start using the skill right away by visiting behindthecraft.com to unlock it and 10 other handcrafted skills.

Upgrade to Get Instant Access

2. Keep reading to understand how I built the skill in 6 steps.

How I built the skill in 6 steps

Here's how the skill works under the hood in 6 steps:

1. Read skill.md and styles.md to understand the guidelines and slide formats.

2. Clarify the topic by asking the user questions about what the contents should be.

3. Research online to supplement what the user provided.

4. Generate the deck as a single HTML file.

5. Run a visual QA loop to take screenshots of each slide and fix layout issues.

6. Deliver the final deck that you can present in your browser.

To build this skill yourself, follow my tutorial and iterate with AI on the steps above.

With AI, you have to build the system first

If you remember just one thing from this tutorial, make it this:

With AI, build the system once instead of doing the work every time.

Creating my /slides skill took me longer than making a deck from scratch. But now that it’s up and running, I can use it to save time on every deck I make from here on out. The same idea applies to any workflow or routine you do more than once.

Watch my tutorial now and let me know if you have any questions in the comments or get a Behind the Craft membership to unlock the skill now.

Today, I want to share a new episode with Ryan Carson.

Ryan is a serial founder who’s building his latest startup solo with only AI agents. It was fascinating to watch him demo the exact skills and routines he uses to get OpenClaw, Codex, and Devin to manage his inbox and calendar, do sales outreach, and build features while he sleeps.

Watch now on YouTube, Apple, and Spotify.

Ryan and I talked about:

• (00:00) Agents are just cron jobs and text markdown files

• (02:25) Demo: How Ryan set up OpenClaw to be his AI chief of staff

• (08:14) The setting most people get wrong in OpenClaw

• (13:46) How to onboard your agent like you would an employee

• (17:06) The cron job that books cold meetings while Ryan sleeps

• (24:36) Why Ryan prefers Devin over other AI coding tools

• (28:33) Demo: How Ryan ships 10 PRs a day solo

• (33:08) Google Ads agencies and a $6K/month designer

• (37:11) Why building the system is the real startup work

I’m proud to partner with Wispr Flow

There are many AI voice dictation tools out there but I’ve found Wispr Flow to be the most accurate at transcribing words accurately. My kids love using it to dictate to Codex and Claude Code to build games and apps. Try it out with my free link below.

Try Wispr Flow for Free

Top 10 takeaways I learned from this episode

Build your AI chief of staff with OpenClaw

1. Make OpenClaw your AI chief of staff with these 4 skills. Each skill is a markdown file that teaches OpenClaw how to do one job:

   1. executive-assistant: Checks email every 15 minutes, watches your calendar, and pings you on Slack when there’s an important email or calendar conflict.

   2. business-development: Searches for sales prospects nightly, adds them to a Google Sheet, and sends personalized email outreach as your agent.

   3. daily-task-manager: Keeps one canonical task list instead of guessing from chat history.

   4. daily-task-prep: Preps your task list before you wake up, pulling from calendar, unread emails, and outstanding outreach.

To copy these skills, first install OpenClaw (or Hermes). Then paste this link into Codex or Claude Code to duplicate Ryan’s full chief of staff system:

Read more

Here’s what you get when you sign up:

1. 10 hand-crafted AI skills that I rely on for personal (Personal Advisor, Health Coach, Weekend Planner), product (Strategy Writer, Spec Writer, Exec Reviewer), and creator work (Social Writer, Infographic Designer, No AI Slop).

2. 3-6 months free for Wispr Flow, Granola, Replit, Linear, and more ($270 value) for new paid users of each AI tool. Plus 20+ prompts that I use personally.

3. An invite to my member-only Slack community with 700+ builders.

Upgrade Now to Get Instant Access

• Already a paid subscriber? Sign in with your paid subscriber email.

• Not yet a subscriber? Upgrade now and you’ll get an invite link in your email.

Watch this 8 min video to see me give a tour of behindthecraft.com:

I plan to add full guides for OpenClaw, Hermes, Claude Code, and Codex — the 4 products that have changed how I work. Prices will go up once all of these guides are live, so subscribe to lock in the early bird price.

10 handcrafted AI skills that you can set up now

Most AI skill packs dump 100 skills on you and let you sort the slop yourself. The skills below are what I actually use every week and all 10 are available now in the Behind the Craft portal.

Personal

1. Personal Advisor: My favorite skill. Get honest advice grounded in your goals, energy, and life situation.

2. Health Coach: Get evidence-based tips on workouts, meals, and sleep, based on your health metrics and routines.

3. Weekend Planner: Find 10 time-sensitive things to do on an upcoming weekend personalized for your location, preferences, and constraints.

Product

4. Strategy Writer: Shape a crisp one-pager document with vision, principles, strategy, and milestones.

5. Spec Writer: Turn rough notes into a clear spec with my personal PRD template.

6. Exec Reviewer: Set up a profile for your CEO or leader and get their feedback before the real review.

Creator

7. Social Writer: Turn your draft into 3 LinkedIn/X posts based on my most viral hits.

8. Infographic Designer: Create a simple design system before generating beautiful, consistent infographics for your content.

9. No AI Slop: Removes 10+ types of AI slop from your writing.

I’ve also included Skill Editor, the skill I use to build and edit every other skill. More on that below.

Upgrade Now to Get All 10 Skills

How I created these AI skills

Skills are the best way to share your expertise with others, so I treat building one like creating a real product. Specifically, I go through 5 phases:

1. List the most repetitive, manual tasks. The ones I do by hand or that require a lot of copying and pasting into AI are the best candidates. For example, I often ask AI for business advice (Personal Advisor) or to remove AI slop (No AI Slop).

2. Separate the instructions from the personal context. I split each skill's instructions (skill.md) and my personal context (plan.md) into two files. The context file auto-updates from past chats so the skill improves over time.

3. Get the AI to ask questions to gather context. Many skills tell the AI to ask a few questions first. For example, the first time you run Weekend Planner it’ll ask you about your location, preferences, and constraints.

4. Spin up sub-agents to test the skill and run evals. A set of agents test the skill with common and adversarial queries, while another set grades each response pass/fail. The AI keeps iterating until everything passes.

5. Test and polish the skill by hand. I always test each skill and make manual edits to improve it. This is the difference between slop and a skill that’s worthwhile.

I’ve found that these steps lead to much better skills. You can use the Skill Editor skill in the portal to follow the same process to create your own.

What’s coming next

I’ve been creating content for 3 years, and it’s only because of my paid subscribers that I can keep doing this. Over the next few months, I plan to deliver to them:

1. Builder skills, starting with skills to create beautiful slides and landing pages.

2. My full personal OS for Hermes / OpenClaw and Codex / Claude Code.

Visit behindthecraft.com to unlock all my skills and prompts plus $270 in free AI tools now.

Upgrade Now to Get Instant Access

Today, I want to share a new episode with Alex Albert.

Alex is a research PM at Anthropic building the next Claude model. He gave me a rare inside look at how the research team operates — including how they turn user feedback into model training, how they decide which capabilities to prioritize, and how they improve Claude’s personality. If you’re wondering how frontier model development actually works, then this interview is a must-watch.

Watch now on YouTube, Apple, and Spotify.

Alex and I talked about:

• (00:00) How Anthropic treats each new model as a product

• (04:58) Building adaptive thinking into Claude

• (07:07) Why Claude is starting to dream

• (11:15) "If it's not a one-way door, then it's essentially free"

• (15:40) How Alex uses Claude Cowork to pressure test his docs

• (17:05) Inside Anthropic's eval process for new models

• (21:02) How Anthropic trains Claude's character

• (32:39) The consciousness question Anthropic is quietly working on

I’m proud to partner with Oceans

I have to give a shout out to Char, my executive assistant from Oceans, for helping me run the podcast. Here’s how we work together:

1. I gave him my AI skills for drafting show notes and clips

2. He’s iterated on them across episodes to make them better

3. We both do a manual review at the end to apply human taste

Now we have a system that saves us hours every week. If you need help with marketing, operations, finance, or an executive assistant, I highly recommend giving Oceans a try. Book a call at oceanstalent.com/peter.

Get Started for Free

Top 10 takeaways I learned from this episode

I’ve organized these takeaways into 3 sections:

1. How Anthropic is building the next Claude

2. How to get your AI product ready for the next model

3. How Anthropic research PMs work

How Anthropic is building the next Claude

1. Anthropic decides what each model should be good at before training. Alex’s team picks the big capability bets upfront before pre-training (e.g., coding, knowledge work). These bets come from talking to enterprise customers and from Anthropic employees hitting roadblocks in their own work.

2. Think about the model and the harness together. The model and the harness are coupled. Each surface wraps the model in a different prompt and tool setup, so the same model can give different responses depending on where it runs. As a research PM, Alex has to think through how the model will perform across Claude, Cowork, Claude Code, and more.

3. Anthropic has full-time researchers thinking about Claude’s consciousness. Anthropic has people whose whole job is to think about what it means for Claude to be a conscious actor. There’s no official position on whether it is or isn’t, but the question is taken seriously as agents take on more autonomous work.

How to get your AI product ready for the next model

4. Test on the latest model and build minimal scaffolding. The two most common mistakes that Alex sees teams make when building AI products:

Read more

Tell me if you’ve seen this before:

Today is a hard day. We’ve made the difficult decision to reduce our workforce by 1,000+ employees. Our business has never been stronger, but AI has changed how we work, and we’re restructuring to win in the agentic era...

Whenever I read one of these, I mentally translate it to what’s actually happening:

Today is a hard day. We screwed up and overhired during the zero-interest era. Now, we need to cut costs. AI agents are easy to blame, so…

According to levels.fyi, over 80,000 tech employees have been laid off in Q1, the highest level since the 2022-23 recession. Here’s a rant about why layoffs continue to happen and what you can do to take back control.

1. The “AI made us do it” copy-paste memo is an excuse

Every layoff memo looks the same these days:

1. “The way we work has changed...Today’s actions are not a cost-cutting exercise; they are about defining how a growth company operates in the agentic AI era.”

2. “This generation of AI is the most transformative technology we’ve seen...We need to be organized more leanly, with fewer layers and more ownership.”

3. “AI tools have changed what it means to build and run a company. A significantly smaller team, using the tools we’re building, can do more and do it better.”

The pattern is always the same: "Today is a hard day → but the business is strong → AI changes everything." It's almost as if the CEOs writing them have a group chat where they share the same script. I think they should just admit that…

2. The real reason is overhiring and empire building

Here’s how headcount grew in the ZIRP era at companies that did Q1 layoffs:

1. Meta: 45,000 in 2019 → 86,000 in 2022

2. Amazon: 341,000 in 2016 → 1.6M in 2021

3. Salesforce: 49,000 in 2020 → 80,000 in 2023

4. Block: 3,800 in 2019 → over 12,000 at peak

Let’s be real — the leaders who grew headcount this quickly messed up. Jack Dorsey, to his credit, basically admitted this the day after cutting 40% of Block’s staff:

There are two main drivers of organizational bloat:

1. Companies prioritized growth at all costs. Money was free and companies chased users and ARR instead of retention and profitability.

2. Empire building was the norm. Leaders were rewarded for growing their scope and headcount instead of keeping their team lean.

Growing headcount this fast is a one-way door change because…

3. The alignment tax is real

If you've ever worked inside big tech, you’ve probably experienced something like this:

You have a product idea that should ship in a week. You just have to draft a memo or deck to get approval from stakeholders and your management chain. By the time you finish the process, months have passed and the market has moved on.

I've worked at great companies with slogans like "it's still day one" and "move fast and break things." But no matter how catchy the slogan is, I now believe that:

As a result…

4. There are basically two types of companies now

If you're a tech employee and worried about keeping your job, start by asking yourself:

Type 1 companies grow headcount as a last resort, even though business is doing well

For example, Anthropic is hiring a lot of people, but only because revenue grew 80x in Q1 2026. When I asked Boris (creator of Claude Code) how his team thinks about hiring, he said he only wants to hire if they've genuinely hit a ceiling on what work AI can take on.

Another example is Linear. Karri (CEO) set a rule to at most double the team every year, even though they've been profitable since 2021:

Type 2 companies are growing headcount despite a stagnating business

You can tell if you’re working in one of these companies if the hallway conversations include a mix of:

• “I need more headcount to do X.”

• “I need to be a manager to grow my career.”

• “Why are we still hiring?”

Unfortunately, I think many tech employees are at type 2 companies where the traditional career ladder is breaking at the seams. From my friend Elena:

Elena’s point is that climbing to VP isn’t the flex it used to be. Smart operators are choosing to go back to IC work instead.

The next generation of great companies have learned from the mistakes of the previous generation. They’re rewarding builders over managers because they know that one person with AI leverage can avoid the alignment tax and do an entire department’s work.

5. How you can take back control as an employee

As an employee, I think the bottom line is:

Take whatever you’re most reliable at and figure out how to do it faster with Claude Code, Codex, or another AI tool that fits your workflow. The goal is to free up your time to do work that’s higher-leverage and harder to commoditize.

I hate seeing employees be at the whim of their employer, so here's what I recommend:

1. Read the signals. If business growth is stagnant and leadership is suddenly obsessed with "flatter orgs," "AI-native teams," or "restructuring for the agentic era," you already know what's coming.

Read more

Today, I want to share a new episode with Moritz Kremb.

Moritz is an AI creator who built a personal OS in Claude Code to manage his emails, create content, and even buy groceries. In this episode, he walks through everything you need to set up your own: folder structure, tools, memory loops, skills, and routines. We also talked about the pros and cons of OpenClaw vs. Claude Code.

Watch now on YouTube, Apple, and Spotify.

Moritz and I talked about:

• (00:00) The honest tradeoffs: OpenClaw vs. Claude Code

• (09:03) The reliability problem that made Moritz switch

• (15:40) The folder structure that runs Moritz's entire OS

• (20:03) The memory loop that makes Claude Code remember everything

• (23:47) CLI vs. MCP vs. API: which to pick and why it matters

• (26:02) Skills walkthrough: automated grocery runs and video uploads

• (29:03) Local vs. remote routines (and when to use each)

• (31:51) The full content pipeline from idea to posted video

• (41:43) Tips to start building your first personal OS

I’m proud to partner with Linear

As teams ship faster with AI agents, the bottleneck shifts from building to deciding what to build.

Linear’s new Agent understands your full workspace context: roadmap, issues, customer requests, and code. Ask it to surface patterns across feedback, scope out a spec, or catch you up on team progress.

Try Linear for Free

Top 10 takeaways I learned from this episode

1. Build your Claude Code personal OS in 4 layers. Moritz runs his email, content pipeline, and grocery shopping through his Claude Code OS. It has 4 layers:

   1. Folder (with memory): A set of files Claude Code reads at the start of every session so it knows who you are, what tools it has, and what it remembers from past chats.

   2. Tools: The APIs, MCPs, and CLIs that let Claude complete tasks like send emails, post to social, search Google Drive, and more.

   3. Skills: Detailed instructions for tasks that Moritz does over and over like “upload my video files” or “refill my grocery cart.”

   4. Routines: Scheduled jobs that run proactively.

Let's walk through how to set up each layer below:

2. Set up your folder structure first. Get this right before setting up tools or skills so that your OS stays organized. Critical files include:

Read more

Today, I want to talk about the race to build a personal AI agent.

Everyone wants an AI chief of staff that can triage emails, book meetings, edit documents, and take care of work that we don’t want to do.

I've spent the last few months testing OpenClaw, Hermes, Claude Code, Codex, and Gemini as personal agents. In this post, I'll share:

1. 10 capabilities a great personal agent needs

2. An honest take on where each product stands today

3. The personal agent stack that I use right now

10 capabilities that a personal agent needs

I think a great personal agent needs to be able to:

1. Manage email, calendar, docs, sheets, and slides

2. Integrate with any API, MCP, or CLI

3. Run recurring and triggered tasks

4. Remember things about you

5. Work across web and mobile with no friction

6. Switch between text and voice replies

7. Have a personality that’s fun to talk to

8. Use your computer and browser

9. Stay reliable

10. Keep your data safe

The reality is that no product checks all 10 boxes today. Here’s how each stacks up:

OpenClaw: The most flexible personal agent, but reliability is a real problem

As the original personal agent, OpenClaw still gets a lot right:

1. Lives inside your messaging apps. Talking to my OpenClaw in Telegram just feels more personal than talking to another agent in a separate app or terminal.

2. Easily switches between text and voice. I primarily talk to my OpenClaw using voice replies, which is convenient and fast.

3. Incredibly flexible. You can customize OpenClaw’s memory, capabilities, and integrations more deeply than any other option in this post.

But OpenClaw’s critical flaw is reliability. I estimate that 10% of my time with OpenClaw is spent fixing it instead of using it. Examples:

1. It forgot it had access to edit Google Docs.

2. It randomly started using a robot voice instead of the one I like.

3. It breaks half the time after every update.

Having to use Claude Code or Codex to fix my OpenClaw doesn’t feel great.

My verdict: OpenClaw is still the most powerful and flexible personal agent. But the maintenance tax is real and reliability is a major concern.

Hermes: I was skeptical at first, but it breaks less often than OpenClaw

I’ll be honest, I avoided Hermes for a long time. Too many people were shilling it on X, and the luxury handbag name felt off. What changed my mind was watching AI builders I trust quietly migrate over from OpenClaw.

I’ve been testing Hermes for the past week. Early impressions:

1. It’s more reliable than OpenClaw. Hermes works more consistently. It even fixed some of the cron jobs that OpenClaw on GPT 5.5 broke:

   

2. It communicates more what it’s doing. It tells me in Telegram when it completes a cron job and how long a task is taking.

3. It tends to work more independently. When it spots a repetitive workflow, it turns it into a reusable skill automatically. This can be a pro or a con depending on how much control you want to give up.

OpenClaw still feels more “alive” thanks to its transparent heartbeat and memory system, but I’d take reliability over those features any day.

My verdict: If OpenClaw’s maintenance tax is wearing you down, give Hermes a try. A week in, it’s been more reliable for me.

Claude Code: The best personality, but availability and rate limits are a concern

Over the past few months, Anthropic has shipped many features to replicate OpenClaw’s functionality, including routines, remote control, and chat channels. As a personal agent, Claude Code has real strengths:

1. Best model personality. Opus feels the most like talking to a trusted friend who will support and challenge you.

2. Strong agentic capabilities. Opus is also, along with GPT 5.5, the best model at just figuring things out, no matter how wild the request.

3. Claude Code is a rabbit hole. Using Claude Code feels like playing a video game where you’re always discovering new shortcuts. It’s a rabbit hole in the best way.

But Claude Code also has major reliability issues:

1. 98% uptime isn’t great. It feels like Claude is randomly unavailable every other week. Key features like routines also break silently without sending any alerts.

2. Mobile integration is manual. You have to type /remote-control to continue your chats on mobile, and it sometimes disconnects after a while.

3. Rate limits are stricter. You can’t get as much usage out of the equivalent Claude plans as you can from Codex (the next option).

My verdict: Claude has a great personality and strong agentic capabilities. But Anthropic needs to prioritize fixing reliability and scaling compute.

Codex: An incredible desktop app and generous limits, but no mobile yet

As a long-time Claude subscriber, I’m genuinely impressed by how much progress the Codex team has made. They’ve shipped:

1. A beautiful desktop app. The app is simple and intuitive to use. You can hook it up to other popular apps with a few simple clicks instead of having to remember slash commands:

   

2. GPT 5.5 with generous rate limits. GPT 5.5 is a great model and you can also get more usage out of the Codex plan than the equivalent Claude plan.

3. Best-in-class browser and computer use. GPT’s strengths here help your personal agent complete workflows in products that don’t have great API support.

Codex has one big gap: you can’t talk to it on mobile yet. 80% of my conversations with my personal agents happen on mobile, so this is a deal breaker.

My verdict: Codex has won me over for coding. It has the potential to be a great personal agent once the mobile app ships (probably very soon).

Gemini: This is Google’s race to lose

Gemini is best positioned to be the go-to personal agent for the masses, but it just doesn’t have all the capabilities yet. What Gemini does well:

1. Native Workspace access. The Gemini app reads emails, schedules meetings, and creates docs by default. But it can’t edit docs yet (more on this below).

2. Best live voice and video conversations. Gemini Live lets you talk back and forth and even supports camera and screen sharing.

But the Gemini app still can’t edit Google Docs, Sheets, or Slides. All of these products are owned by Google, so it’s not great that Gemini is the only AI app that doesn’t support this. Codex and Claude Code do it natively with a few simple connections.

Overall, Google seems to be adding chat windows to all of its individual products instead of making the Gemini app the most capable personal agent.

My verdict: Google has all the data and products to win this race. But basic capabilities are still missing, and time is running out.

My personal agent stack right now

As of this writing, I use:

1. Hermes for everyday tasks. Emails, calendar, Google Doc edits, and honestly, therapy through voice replies. I switched over to Hermes because it’s more reliable than OpenClaw.

2. Codex and Claude Code for building things. When I’m coding, writing, or producing real work, I live in these apps. If I had to pick one, it’d be Codex due to the more generous rate limits (but you know how these things change fast).

The important thing is to not get stuck like me, migrating from one personal agent to another instead of doing real work. Pick one or two agents that work for you based on the pros and cons above and just commit.

One thing I can promise you:

Once you have an agent that’s available 24/7 and can actually get work done for you, you’ll never go back to a regular AI chat interface again.

Today, I want to share a new episode with Ravi Mehta.

Ravi is a top AI instructor and former CPO of Tinder. In our episode, he did a live demo of using his 3-layer context engineering system (functional, visual, data) to build a beautiful music app instead of the usual AI slop. You’ll never prompt AI the same way again after learning about Ravi’s approach.

Watch now on YouTube, Apple, and Spotify.

Ravi and I talked about:

• (00:00) Why most AI prototypes feel like slop

• (01:08) Context engineering, explained from first principles

• (05:12) Layer 1 demo: Functional context

• (09:38) Layer 2 demo: Visual context

• (13:31) Layer 3 demo: Data context

• (15:47) Building a custom MCP server in Claude Code

• (19:54) The full stack prompt: all 3 layers in one shot

• (24:21) Why separating the data layer is the real unlock

• (28:59) Should PMs prototype or just ship to production?

• (33:51) Where PRDs still fit in AI native product development

I’m proud to partner with Wispr Flow

Wispr Flow has quietly become one of the AI tools that I use the most. It lets you dictate into any app on your computer or phone and turns your voice into clean, formatted text automatically.

I save at least 3 hours a week from using it to draft newsletter posts, write specs, and reply on Slack. It’s just so much faster to dictate to AI than to type, and you’ll never go back once you try it. Use code PETERWISPRFLOW to get 6 months free.

Try Wispr Flow for Free

Top 10 takeaways I learned from this episode

1. Apply three layers of context to create polished AI prototypes. Ravi built the same music app in 3 ways to highlight what each context layer adds:

   • Functional context. First, he wrote a spec listing the feature requirements (hero, albums, etc) and got a generic app with placeholders.

   • Visual context. Next, he added a wireframe from Figma to the prompt. This led to better UX and layout but the product was still missing real data.

   • Data context. Finally, he asked AI to generate a JSON with genre, artist, and album names and used a custom MCP server to pull in album cover art from Deezer. This time, AI generated a complete music discovery app.

The best thing about this system is that you can swap the data layer (JSON) at any time to have your app highlight completely different content (e.g., from rock to hip hop albums) without touching the other two layers.

Here’s a detailed prompt that Ravi shared with me to generate all 3 layers in one-shot for any app or idea that you have (just replace the music app details with yours):

Read more

OpenAI vs. Anthropic gets all the attention these days, but Google Labs is quietly shipping some of the most interesting AI products right now.

Pomelli for marketing, Stitch for design, Flow for videos, and more — these tools are better than most people realize.

Watch the tutorial now to see me review all five tools live and share where each product excels and where they fall short.

Timestamps:

• (0:00) Why Google’s most underrated AI products are in Labs

• (0:39) Pomelli: Generate marketing assets from any URL

• (4:37) Stitch: Create web and mobile designs from scratch

• (7:27) Genie: Walk through AI-generated 3D worlds

• (9:11) Flow: The best tool for making AI videos and images

• (11:49) NotebookLM: Turn any source into video, slides, and infographics

• (14:27) 3 takeaways on Google’s AI strategy

I’m proud to partner with Granola

The world’s best product teams including Ramp, Linear, Replit, and more all use Granola to take notes and focus more in meetings. I’ve tried many meeting notes apps myself but keep coming back to Granola. It’s just that good.

Try Granola for Free

A review of 5 Google Labs tools and their pros/cons

1. Pomelli: Generate marketing assets from a link

Pomelli generates beautiful marketing assets and your brand DNA (colors, fonts, images, logo), all from a link to your website.

For example, I gave Pomelli a link to my favorite local Asian bakery, Simply Cake. It pulled the brand’s logo, fonts, colors, and images in under 20 seconds:

Once Pomelli has your brand’s DNA, it’ll generate 3 marketing campaigns for you to review. Here’s a Mother’s Day campaign it made for Simply Cake:

Pomelli’s Photoshoot feature is also worth trying. Upload a basic photo of an egg tart and it creates marketing-ready assets with Nano Banana:

Next, I tried Pomelli on my newsletter to see how it handles a content business. Unfortunately, it didn’t work as well. Content brands like mine don't have great product photography to extract from, so Pomelli has less to work with:

My verdict: Pomelli is incredible if you run a physical or e-commerce store with a strong visual library. My one gripe is that you can’t edit the text and images manually — you have to re-prompt the AI every time.

2. Stitch: Create beautiful product designs

Stitch is Google's answer to Claude Design. I gave it a simple prompt to "design a mobile fitness app." Stitch first generated a design system, then produced the screens:

One great thing about Stitch is that it’ll use Nano Banana to fill in images, which makes the screens look more realistic. There’s also a fun feature that shows a predictive heat map of where users might click.

But Stitch has one flaw compared to Claude Design — it didn’t ask me a single clarifying question before making the app. Compare that to how Claude Design handles the same prompt:

Read more

Today, I want to share a new episode with Tibo Louis-Lucas.

Tibo is a solo founder making $1M+ a month (!) after bootstrapping 5 AI products. I got him to walk through his exact playbook for validating ideas fast, acquiring users, and pricing AI products. If you want to bootstrap an AI business from scratch, this interview is a must watch.

Watch now on YouTube, Apple, and Spotify.

Tibo and I talked about:

• (00:00) The right mindset for solo AI founders

• (04:48) Tibo's 5-rule playbook to reach $1M/month

• (09:55) The product pivot that unlocked $600K/month

• (13:08) Why you should charge money right away

• (20:50) The acquisition channel everyone says is dead (it's not)

• (26:58) Running 5 businesses without a real team

• (36:34) How to price your AI product

• (37:20) What's a good churn rate for AI products?

• (40:16) 9 products failed. The 10th hit $600K/month

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 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

Top 10 takeaways I learned from this episode

1. Tibo’s 5 rules behind his $1M/month AI portfolio. After launching 5 successful AI products (and many failed ones), Here’s Tibo’s rules to build an AI product that customers want:

   • Charge money from day one. Collect money early so you know if people actually want what you're building.

   • Follow user behavior, not your plan. When people twist your product into something unexpected, build for that instead.

   • Price for quality customers. Charging $50-100/month keeps away high-churn users who cancel fast and flood customer support.

   • Build free tools that rank on Google. Target each tool to a specific search query (e.g., “audio to video”) to win SEO traffic.

   • Launch fast and kill what doesn't work. Accept that most ideas will fail, so move quickly and double down on what sticks.

Let's dive into these rules with real examples 👇

Read more

Today, I want to share a new episode with Ryan Wiggins.

Ryan is the VP of Product at Mercury, my favorite fintech product. We covered three things that are on every AI builder’s mind — how to make your product agent-ready, how to build a Claude Code second brain, and what Mercury’s data reveals about OpenAI and Anthropic’s race for the enterprise.

Watch now on YouTube, Apple, and Spotify.

Ryan and I talked about:

• (00:00) Why APIs and MCPs are the primary interface

• (04:40) How to get your product agent ready

• (07:51) Will MCPs cannibalize your app DAU?

• (10:36) How to decide between MCPs and CLIs

• (13:10) Inside Ryan's Claude Code second brain

• (18:40) Getting AI to coach you after every meeting

• (20:46) Mercury data: Are startups switching to Anthropic?

• (23:05) The product development process will never be the same

P.S. Mercury is offering Behind the Craft readers 1 year of free personal banking. Just click the link to get the discount.

I’m proud to partner with Wispr Flow

For many PMs and tech professionals, communication is the job. Slack replies, stakeholder updates, PRDs, emails — most of us are bottlenecked by typing speed.

Wispr Flow is the tool I use to close that gap. You speak naturally into any app and it outputs clean, ready-to-send text. It genuinely saves me 3 hours per week and works across Mac, Windows, iPhone, and Android. Use code PETERWISPRFLOW for 6 months free.

Try Wispr Flow for Free

Top 10 takeaways I learned from this episode

How Ryan built his Claude Code second brain

1. Ryan built a Claude Code “second brain” trained on 5 years of work history that helped him 2x his productivity as a product leader. He did this in 5 phases:

   1. Create a profile. Have Claude interview you and write a me.md file capturing your role, goals, working style, and growth edges. This is the foundation.

   2. Build the knowledge base. Save your most important docs — specs, retros, performance reviews — in a local folder. Then ask Claude Code to install QMD (Shopify CEO’s vector search tool) so it can reference these docs.

   3. Distill the data. Ask Claude to create summary files using me.md and the knowledge base. For example:

      1. strategic-context.md — what my company/team is trying to do and why

      2. personal-growth.md — patterns in my feedback, coaching themes

   4. Wire up auto-context injection. Create a Claude Code hook that pulls relevant context from the files above into every prompt.

   5. Create the learning loop. Build a /learn skill that updates Claude.md after each session, plus cron jobs for morning briefs and monthly retros.

Ryan uses this to drive team priorities, make decisions, and self-reflect. It’s the most sophisticated personal OS I’ve seen, so I asked Ryan for a full 4,000+ word prompt that you can paste into Claude to set this up for yourself in 5 steps:

I want you to help me build a "Second Brain" — a persistent knowledge system that runs in parallel to my work using Claude Code. We'll do this in 5 phases. Walk me through each one interactively. Don't skip ahead — confirm each phase is working before moving on.

## Phase 1: Interview Me & Create my Profile

Interview me to create a file at ~/Documents/second-brain/me.md that captures:
- Who I am (role, company, responsibilities)
- What I'm optimizing for (goals, priorities, what success looks like)
- My working style (tools I use daily, how I communicate, what frustrates me)
- My growth edges (feedback I've gotten, patterns I want to break)
- What I care about outside work (interests, values — helps with recommendations)

Ask me 5-7 questions conversationally. Don't make me fill out a template. Write the file when you have enough.

## Phase 2: Build the Knowledge Base

Read more

Today, I want to share a review of Claude Design, Anthropic’s new AI design tool.

In the tutorial below, I ran through 5 real demos live using Claude Design to create videos, slides, websites, apps, and even a design system.

Watch my tutoral now for the complete walkthrough.

Timestamps:

• (0:00) Why Anthropic is coming for the full knowledge work stack

• (1:12) Turn any post into an animated video

• (4:17) Pro tip for creating beautiful slide decks

• (7:27) Recreate any landing page in one prompt

• (9:21) Build a 3D rotating globe and text particle effects

• (10:53) One-shot a mobile fitness app with play testing built in

• (12:46) Generate an Apple Liquid Glass design system

• (14:28) 3 takeaways on where AI design is headed

I’m proud to partner with Granola

I have Granola’s MCP connected to Claude and use it as a weekly coach. Every Friday, I ask it to recap my meetings and pull out next steps to follow up on.

The better prompts are the reflective ones: “What themes keep coming up across my 1:1s?” or “Where did I talk too much and listen too little?” Your meeting notes know you better than you do. Granola makes that context easy to use.

Try Granola for Free

Anthropic is eating the knowledge worker stack

Anthropic is quietly building an AI native version of the entire knowledge worker stack. They have:

1. Claude Code for coding

2. Claude Design for design

3. Claude Cowork for documents and more

The common thread across all three products is simple:

Code is the universal medium for all knowledge work.

Videos, slides, websites, apps, and animations — they’re all just code under the hood. And code is now free, or at least very cheap.

This means Claude now competes with both incumbents like Figma and Microsoft Office and AI startups like Replit, Lovable, and Gamma.

But how good is Claude Design? Let’s run through 5 real use cases: animated videos, pitch decks, landing pages, mobile apps, and design systems.

1. Animated video

Let’s start with video. Open a new project in Claude Design and paste in this prompt:

Make an animated video for Claude Design featuring this content: (Paste in the Claude Design announcement post or any text you want)

Claude will ask you clarifying questions first — how long the video should be, what tone and aspect ratio you want, and more — before making the video. It’s the same behavior as Claude Code when you ask it to build something:

Here’s the video Claude Design made in one shot:

This would’ve taken hours or days to do manually but Claude did it in just a few minutes. Because the video is just code, you can also manually tweak the color palette and even turn captions on or off.

The only flaw is that you can’t export the file as a video yet, so I had to use Mac’s screen recording to save it.

2. Slide deck

Now, let’s try using Claude Design to make a slide deck. When using AI to make a deck, I like to use regular Claude to iterate on the copy for each slide first:

Next, I went to Claude Design, selected Slide Deck, and prompted it:

Make a slide deck with this copy: (TEXT)

Here’s an example slide that it made:

This looks OK but the overall deck feels too static and boring compared to the video we just made. So my pro tip for creating beautiful animated slides is:

Read more

Today, I want to share 15 observations from my first trip to China in 10 years.

I just came back from a 2-week trip to Shanghai and Hangzhou and it was eye-opening to see how the world’s 2nd largest economy operates today.

You can’t understand the AI race without understanding China and I think every product builder should visit at least once. I hope this post gives you a taste of what it’s like to live and work there.

I’m proud to partner with…Wispr Flow

Wispr Flow has quietly become one of the AI tools that I use the most. It lets you dictate into any app on your computer or phone and turns your voice into clean, formatted text automatically.

I save at least 3 hours a week from using it to draft newsletter posts, write specs, and reply on Slack. It’s just so much faster to dictate to AI than to type, and you’ll never go back once you try it. Use code PETERWISPRFLOW to get 6 months free.

Try Wispr Flow for Free

Chinese AI work culture isn’t for the faint of heart

1. 996 is real, but many tech workers show up and work even later. Employees at many Chinese AI startups roll in around 11 AM and work until 11 PM with lunch, dinner, and boba breaks in between.

2. Top AI labs run on young, single talent. The workforce at top AI companies in China is overwhelmingly people in their 20s and early 30s. They don’t drink, smoke, or party much — many just work and order food and boba delivery to the office. In many ways, this looks a lot like AI work culture in the Bay Area.

3. Beijing, Hangzhou, and Shenzhen are the AI tech hubs. Beijing has Z.ai’s GLM, Moonshot’s Kimi, and ByteDance’s Doubao. Hangzhou has Alibaba’s Qwen and DeepSeek while Shanghai has MiniMax. Shenzhen is the global hub for robotics and physical AI with 57,000+ robotics companies.

4. Many Chinese tech employees rely on Claude and Codex via VPN. Despite Anthropic and OpenAI blocking access in China, employees at Chinese AI labs use VPNs and special credit cards to access the best US models. Bypassing the Great Firewall is routine for China’s younger generation.

5. The government is pouring billions into AI and one-person companies (OPCs). Youth unemployment (age 16-24) sits at 16%+, so the government is heavily encouraging people to start one-person companies. Cities compete for the best AI founders by offering subsidies like Hangzhou’s 1B yuan ($140M USD) OPC fund.

Living in China feels like the future (especially if you’re on US income)

6. Ride sharing and delivery are incredibly cheap. DiDi rides cost just $3-10 USD for most city trips compared to $30+ for an Uber in San Francisco. You can also get almost anything delivered for a few bucks. I got boba delivered to my hotel room for $2 and it arrived in a fancy champagne bag. Many younger families just order delivery for dinner because it’s more convenient than cooking.

7. Live-in nannies and home chefs cost a fraction of US prices. A standard live-in nanny in Shanghai runs $1,100-1,700 a month all-in and a part-time helper for cooking and cleaning is $7-10/hour. In Hangzhou, I met a mom who brought her nanny on vacation just to take care of the kids.

8. WeChat runs payments and everything else. It’s an entire operating system in China with mini apps for ride sharing, public transportation, food delivery, and more. Payment is almost entirely done through QR codes — I didn’t pull out my credit card or cash a single time in two weeks.

9. Chinese electric vehicles have taken over the roads. New EVs hit 54% of all new car sales in 2025, and Chinese brand BYD has overtaken Tesla globally with 2.26M electric cars sold. The BYD Qin L starts at $16,500, which is half the price of the Tesla Model 3 at $33,500.

10. The infrastructure is more modern than New York or Tokyo. Shanghai Metro is 819 km vs. NYC’s 472 km, and China operates 50,000+ km of high-speed rail, or two-thirds of the world’s total. Riding the Shanghai Metro and then reading about Bay Area BART train budget cuts is a strange experience.

The social contract in China is fundamentally different

11. There are cameras everywhere. China has 750-800 million surveillance cameras — more than half the world’s total, or 1 camera per 2 people. Walk down any sidewalk in Shanghai and you’ll see poles with 7-8 cameras on them. Living there, it really does feel like every action is being watched and recorded.

12. As a result, almost nobody breaks the rules. There’s very little speeding or street crime. China’s blood alcohol limit is 0.02% (vs US 0.08%), so people won’t even take a single sip if they’re driving. Walking around at night with my kids felt safer than in most US cities I’ve lived in.

13. Family gatherings revolve around 3-hour meals at high-end malls. Multi-generational families gather at mall restaurants, order way too much food, share a bottle of baijiu, and stay for hours. The mall is the center of weekend social life, not just a place to shop.

14. The elderly are living their best lives while youth face pressure. Go to any park and you’ll see retirees dancing, playing yo-yo, and singing in groups. Meanwhile, many young people are “lying flat” — living off their parents’ income with no plans to chase a career.

15. The post-2022 expat community is much smaller and more committed. The 2022 Shanghai COVID lockdown broke a lot of trust among expats. Combined with geopolitical tightening and a slower economy, the casual “3 years and out” expat crowd is mostly gone. What’s left is a smaller community of people married to locals or running businesses there.

5 tips if you’re thinking of visiting

China is arguably the most modern country that I’ve been to, so I think every product builder should visit it at least once. But thanks to the Great Firewall and other restrictions, you have to really prep ahead of time.

5 practical tips if you’re thinking of making the trip:

Read more

Today, I want to share a new episode with Dylan Field.

Dylan is the co-founder and CEO of Figma. I asked him tough questions on whether AI can learn design taste, whether design systems constrain creativity, and what design’s role is when code is free. Two-thirds of Figma users are now non-designers and it was super interesting to hear Dylan talk about where he’s taking Figma next.

Watch now on YouTube, Apple, and Spotify.

Dylan and I talked about:

• (00:00) Taste, craft, and point of view: the 3 skills AI can’t replace

• (05:33) Can AI agents learn design taste?

• (15:00) What’s Figma’s role when code is free?

• (17:42) What vibe coding gets wrong about design

• (22:17) Two-thirds of Figma users are now non-designers

• (26:47) Do design systems constrain creativity?

• (29:13) The one trait Dylan hires for over AI fluency

• (32:01) Why Dylan thinks design is the new code

• (36:32) Dylan’s personal AI workflow for writing and ideation

I’m proud to partner with Linear

As teams ship faster with AI agents, the bottleneck shifts from building to deciding what to build.

Linear’s new Agent understands your full workspace context: roadmap, issues, customer requests, and code. Ask it to surface patterns across feedback, scope out a spec, or catch you up on team progress.

Try Linear for Free

Top 10 takeaways I learned from this episode

1. Taste, craft, and point of view are three different things. Here’s how Dylan defines each:

   • Taste is knowing what’s good and being able to articulate why. It’s navigating possibilities with clear preferences.

   • Craft is “pushing past where others might push” at every level, from the macro structure down to the smallest detail.

   • Point of view is the insight that you see, expressed through what you build. Dylan’s test: “If everyone agrees with your POV, you probably don’t have one.”

2. AI gets you to average fast and your job is to push beyond that. “The first thing AI gives you is generic by definition because it’s the average of everything it’s seen.” You need to apply your taste and judgment to iterate with AI toward something only you could have made.

Read more

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

Today, I want to share a new episode with OpenAI’s Codex team.

Alex is the product lead for Codex and Romain runs developer experience. They gave me a rare look at how their team operates including how they build products with Codex and ship without traditional specs or roadmaps. Alex also has some spicy takes on the future of PM and what he looks for when hiring (it’s not resumes!).

Watch now on YouTube, Apple, and Spotify.

We talked about:

• (00:00) Live demo: Building a game in seconds with Codex Spark

• (04:29) "For specs, we write like 10 bullets and that's it"

• (08:09) "Our designers now write more code than eng 6 months ago"

• (15:30) "We plan short and long term, never medium term"

• (22:11) How Alex actually spends his day as Codex’s product lead

• (32:06) The traditional career ladder no longer makes sense

• (39:20) What the Codex team looks for when hiring (it's not your resume)

I’m proud to partner with Linear

As AI agents like Codex become part of your company’s daily workflow, having visibility into what they’re doing is critical.

With Linear, you can delegate issues to Codex directly by simply tagging @Codex in an issue. You can also use the new Linear plugin inside Codex to ensure that your issues and project updates stay synced in real time.

Try Linear for Free

Top 10 takeaways I learned from this episode

How the Codex team ships products

1. The team designed the Codex app to let you delegate to multiple AI agents. The team knew that multi-agent delegation was the future, but felt that having 18 terminals wasn’t the right UX. So they started prototyping apps early before shipping on a beautifully simple agent management interface.

2. Plan near-term or long-term, never medium-term. Alex got this advice from an OpenAI researcher: “Plan concrete milestones up to 8 weeks out and have a long-term vision of where models are heading.” Medium-term roadmaps (e.g., 6-12 months) are awkward and always changing because AI coding is evolving so fast.

Read more

Today, I want to show you how to build a beautiful mobile app to track your workouts.

Building a mobile app used to require a team of engineers. But AI tools like Claude Code and Pencil have made it easy for you to ship a beautiful app by yourself even if you don’t know how to code.

Watch my tutorial to see me walk through how to do this in just 16 minutes.

Timestamps:

• (0:00) What we'll build: A beautiful mobile app to track your workouts

• (2:01) Step 1: Define requirements with Claude

• (6:11) Step 2: Create designs with Pencil

• (8:56) Step 3: Build the app with Claude Code

• (14:21) Getting the app on the App Store

• (15:19) Recap of all 3 steps

Watch now on YouTube or read the written guide below.

I’m proud to partner with…Granola

Granola just launched Spaces — a new way for your team to organize and share meeting notes with each other. For example, you can create a “User Calls” space with your interviews to let anyone ask questions about user needs. You can even ask questions directly in Claude, ChatGPT, or Gemini with Granola’s new APIs and MCP.

Try Granola for Free

What we’ll build: A mobile app to track your workouts

We live in a time where it’s easier to vibe code your own mobile app than to find an existing one that you like. So let’s build an app that can:

1. Create workouts by selecting an exercise and then the number of sets and reps.

2. Track workouts live by tapping each exercise set to complete it.

3. Measure progress via a calendar and charts for each exercise.

I’ll walk you through how to build it in 3 steps below.

1. Define the requirements with Claude

Let’s start by pasting this prompt into Claude:

Let’s work together on requirements for a mobile fitness app. I want three screens:

1. Add/edit workouts: Create workouts and add exercises with sets and reps

2. Workout session: Track workouts in real-time, tap to complete sets

3. Calendar: See which days you worked out and view workout details

This is the basics, but let’s work on the spec together. I want you to stay crisp on the requirements, and if you have any questions, ask them in a numbered list.

The most important part of this prompt is

I always include this line when discussing requirements with Claude so that it can ask me questions to make the spec better. Here are a few example questions and answers:

1. Claude: Do you want a rest timer?
   Me: Yes, also let me customize it in settings.

2. Claude: Do you want automatic deload after failed sets?
   Me: No, let’s keep it simple.

After Claude gets the spec to a state that you’re happy with, ask it to:

Include the last line because Pencil (our AI designer) is better at design than Claude. 🙂 Let’s cover that next.

2. Design the mobile screens with Pencil

This is where the magic happens. Here’s how to use Pencil to create beautifully crafted mobile screens instead of AI slop:

Read more

Today, I want to share a new episode with Jenny Wen.

Jenny is Cowork’s design lead. She gave me a rare inside look at how Anthropic operates — including how she uses Cowork to design and build products and the real story behind how Cowork was built. Anthropic is shipping new features every day and it was fascinating to see how they do it.

Watch now on YouTube, Apple, and Spotify.

Jenny and I talked about:

• (00:00) How design work has changed at Anthropic

• (04:30) "My secret is that I actually use Cowork for most things now"

• (05:19) Cowork demo: From messy user feedback into product insights

• (14:41) "Every Monday Cowork gives me 3 vetted product ideas"

• (23:09) The real story behind Cowork's development (10 days?)

• (27:18) Exclusive look: Cowork prototypes that didn't ship

• (34:01) How product planning works at Anthropic

• (38:10) "If you feel like the ground is shifting, it's because it is."

I’m proud to partner with Replit

Replit’s new Agent 4 is the best way for teams to collaborate with AI agents.

1. Creating multiple design variations on an infinite canvas (see above).

2. Building using parallel agents using a kanban-style task board.

3. Deploying to production in one click.

Get 3 months free to Replit by using my code PYANG26 at checkout.

Get Replit for 3 Months Free

Top 10 takeaways I learned from this episode

Jenny shared a lot of great insights so I’ve organized the takeaways into two sections:

1. How Jenny uses Cowork at Anthropic

2. How the Cowork team designs and ships products

How Jenny uses Cowork at Anthropic

1. Set up Cowork to deliver a product priorities deck every week. Jenny demoed using Cowork to summarize user feedback from customer calls and social posts. She then asked Cowork to use that feedback to create a product priorities deck. Finally, she scheduled a workflow to share an updated deck in Slack every Monday for her team to review.

Here’s a 1,500+ character Cowork prompt that I created to do all of the above in one-shot. Please make sure to replace [folder] with your local folder where you’ve pasted user feedback files and replace [product name] with your product name:

Read more