Today, I want to share a new episode with Shreya and Hamel.

Shreya and Hamel teach the industry-leading AI evals course taken by 4,500+ students from OpenAI, Google, and more. I asked them to do a live audit of the evals that I built for my creator skills. They then walked through how anyone can use their free Error Discovery skill in Claude Code to turn feedback into useful evals. If you want a practical, up-to-date primer on AI evaluations, then this episode is for you.

Watch now on YouTube, Apple, and Spotify.

P.S., I asked Shreya and Hamel to give 25% off the next cohort of their AI evals course that’s starting on 9/5. It has 4.7 stars with 900 reviews, which is almost unheard of on Maven. Check it out and get your company to expense it if you can!

Get 25% Off the Course Now

Shreya, Hamel, and I talked about:

• (00:00) Why AI evals are completely different now with the latest models

• (02:30) Demo: Live audit of the evals that I built for my AI skills

• (05:14) Top-down vs. bottom-up evals and why AI sucks at the latter

• (08:16) Spinning up AI agents to grade each eval criterion

• (14:32) Demo: Using Shreya’s AI skill to run evals in Claude

• (23:46) How to review output labels to identify failures

• (34:37) How to turn failures into reusable evals

• (43:19) Where human judgment is still needed

I’m proud to partner with Wispr Flow

When I’m building with Claude Code, I use Wispr Flow to dictate all my prompts instead of typing. It’s more accurate than other AI voice dictation tools and saves me hours every week. Try it out for free with my exclusive link below.

Try Wispr Flow for Free

Top 10 takeaways I learned from this episode

Live audit of the evals for my AI skill

1. There are two types of evals: top-down and bottom-up. When I asked Shreya and Hamel to audit the evals that I built for my podcast production skill, they found strong top-down evals, but almost no bottom-up evals.

   1. Top-down evals are rules that you define upfront. Mine include “Is every takeaway 240-330 characters?” and “Does each give plain, useful advice?”

   2. Bottom-up evals are failures found by comparing AI’s output with my final edits. With these evals, it’s important not to overfit to a few examples.

2. Compare 10–20 past examples before turning a failure into an eval. Hamel and Shreya encouraged me to run my podcast production skill on at least a dozen past interviews to compare AI’s output with my final edits. They shared a free Error Discovery skill that Claude Code and Codex can use to define evals based on this data.

Build useful evals using this free skill in 5 steps

Shreya and Hamel built a free Error Discovery skill that anyone can run in Claude Code or Codex to define useful evals based on your data. The skill walks you through this process:

Below is a link to the skill and a walkthrough of each step:

Read more

Today, I want to share a personal story about my mom that has occupied most of my time over the past few weeks.

Every summer, my wife and I bring our daughters to Vancouver to visit my parents. We’ve shared many happy memories together, and I expected this visit to be the same.

But a week into our trip, a CT scan found a mass in my mom’s chest. She has survived breast cancer three times, and the radiologist suspects it may have returned.

The first three times this happened, my parents carried most of the burden. This time, I’m not going to be a bystander. Instead, I’m making my mom my #1 priority for the rest of this year.

I’m proud to partner with Wispr Flow

Wispr Flow recently launched Notetaker to help you take effective meeting notes. I’ve been testing it out and love how it automatically recognizes speaker names and turns conversations into summaries that I can act on. Check it out for free via my link.

Try Wispr Flow for Free

A brief history of my mom’s battle against cancer

If you know anyone who has cancer, you know that it casts a dark cloud over the whole family. I’ve experienced this cloud three time before:

1. The first time was while I was in high school. My mom went through chemotherapy and lost her hair, but my parents shielded me from most of what was happening. I knew she was sick without understanding how serious it was.

2. The second time was while I was in college. I barely remember it because I was on the opposite coast attending Brown. My dad took care of her treatment.

3. The third time was when I was getting married. By then, my mom’s cancer had progressed to stage 3, but she didn’t tell me because she didn’t want it to overshadow my wedding. I remember her showing up to the wedding wearing a wig.

After that, she had 10 good years free from cancer. She traveled the world with my dad, and her doctors stopped checking on her after 5 years.

But now her cancer is likely back.

AI is helping me navigate the healthcare system

Needless to say, I turned to AI to understand my mom’s situation:

1. I created a ChatGPT project with her scans and medical records, then shared it with my parents so we could work from the same facts.

2. I got AI to explain the reports in plain English, flag what remained uncertain, and prepare questions for her doctors.

3. I used AI to maintain a one-page brief with her history, the latest updates, open questions, and key phone numbers.

I expected AI to be most useful for researching the disease, but honestly, its biggest value has been helping us navigate the healthcare bureaucracy. A few examples:

1. My mom’s primary doctor submitted an oncology referral. AI helped me realize that getting a biopsy required a separate referral.

2. AI identified the hospital department that performs biopsies and helped me pre-fill the request for her primary doctor.

3. AI helped me find who to call, what to ask, and how to follow up on both referrals.

Here’s one of my conversations with AI:

So far, AI’s biggest contribution has been helping me get the right information to the right people without losing time. After almost daily calls to various medical departments, we now have a biopsy booked for later this week and an oncology appointment for later this month.

Strangers have been generous with their time

As useful as AI is, good old-fashioned human connection has been just as important. I’ve been struck by how willing strangers have been to help:

• Maor, the founder of Base44, raised the idea of preserving extra tissue for deeper testing and connected me with people exploring options beyond standard care.

• Rachel, a breast cancer nurse in Israel, reviewed four scans and explained what the biopsy needs to clarify, including which tumor markers should be retested.

• Poornima and Bayli from the Sijbrandij Foundation advised us to pursue the biopsy referral in parallel and and ask about cancellation lists.

• Katy, my oncologist friend, reviewed the scans and encouraged us to reconnect with my mom’s former cancer team.

I added the notes to the shared AI project after every conversation so we could build on everything we had learned.

I’m putting my mom first

When I shared that I was quitting my job, I wrote that you shouldn’t take the people you love for granted. I didn’t expect that lesson to feel so real only a few months later.

The greatest gift of going independent has been being able to choose how I spend my time. After interviewing hundreds of AI builders and leaders, the person I most want to sit down with now is my mom.

My dad needs to travel to Asia, so I booked another flight back to Vancouver to be with her. I plan to ask her about what it was like to cook for her younger brothers as a teenager, move to a new country, and build a life with my dad from nothing.

The business can wait. I want to hear all her stories.

P.S. If you know about the latest breast cancer treatments or have a personal cancer story you’d like to share, I’d love to hear from you (just reply to this email).

Today, I want to share a new episode with Riley Brown.

Riley is a top AI creator and founder of Agent Native. In this episode, he showed me how he uses Codex to run his entire content business from creating engaging thumbnails to writing viral video hooks. Riley also shared his full AI stack including some lesser-known tools to edit videos, draw diagrams, and more.

Watch now on YouTube, Apple, and Spotify.

Riley and I talked about:

• (00:00) How Codex runs Riley’s 1.5M+ follower content business

• (03:16) Demo: Using Codex to create animations and B-roll

• (07:28) Demo: How Codex can reverse-engineer a top YouTube hook

• (13:24) The BRENS framework to turn an idea into a strong hook

• (17:24) Demo: Riley’s AI stack for diagrams, thumbnails, and posts

• (22:01) How Riley creates so much content every week

• (25:41) Demo: Using Codex and Paper to create viral YouTube thumbnails

• (35:01) How to set up AI agents to find content ideas proactively

I’m proud to partner with Google AI Studio

Google AI Studio was my go-to prototyping tool when I was a product manager. You can use it to generate three versions of your product, add visuals with Google’s Nano Banana model, and publish them to a custom URL for customer feedback. It’s an easy, free way to move from prompt to something your team or customers can try.

Try Google AI Studio for Free

Top 10 takeaways I learned from this episode

1. Remix proven thumbnails with Codex and Paper’s AI design tools. Riley creates winning YouTube thumbnails entirely using AI as follows:

   1. Ask Codex to collect thumbnail formats from top YouTube channels in Paper.

   2. Use Paper’s image-generation features to mix his face with a proven format.

2. Riley’s AI stack helps him run his entire YouTube video creation process. In addition to Codex, Riley uses a few lesser-known AI tools to research topics, draft hooks, draw diagrams, create animations and B-roll. Here’s his stack:

   1. Research topics using Codex and Supadata, which can efficiently pull transcripts for top-performing YouTube videos in a few seconds.

   2. Draw diagrams using Wispr Flow and Excalidraw. He dictates directions to the Excalidraw MCP to turn into diagrams that frame his videos.

Read more

Get instant access

Dear subscribers,

I’ve started to get a surprising amount of work done by talking to AI while walking outside. I believe that:

We’re moving from manually using a keyboard, mouse, and laptop to directing agents in the cloud using voice.

In this post, I’ll cover five shifts that can help you get ahead of this change:

1. Voice will become the orchestration layer

2. Personal computers will move to the cloud

3. Products will be built for agents first

4. Most software will become commoditized

5. Trust will determine the winners and losers

I’m proud to partner with Riverside

Recording a podcast is the easy part, but turning each conversation into clips, show notes, and posts requires a ton of work. Riverside makes this easy for me by providing an end-to-end content creation platform for editing, clips, publishing, and more. Try Riverside free for one month with code PeterYang.

Get 1 Month Free for Riverside

1. Voice will become the orchestration layer

I’ve been using tools like Wispr Flow to dictate instructions to AI for the past year. Recently, ChatGPT Voice has taken this to the next level by making voice the orchestration layer for everything I do on a computer.

While out on a walk or in bed, I’ll turn on ChatGPT Voice and just start rambling:

• “Use my chief of staff thread, check any open email loops I should respond to.”

• “List the top 5 PRs for (project), let’s work through them one by one.”

• “I just interviewed (guest), run the /podcast-production skill in a new thread.”

ChatGPT Voice will provide instructions to all these different threads and report back when each task is done. It honestly feels magical compared to tracking everything myself with a keyboard and mouse.

Recently, I taught my 8-year-old how to use ChatGPT. She doesn’t even know how to type but is already building games with AI just by talking to it.

For her and the generation that’ll grow up with AI, voice is already the default interface.

2. Our computers will move to the cloud

You’ve probably seen the meme of people walking around with their laptop lid half open to keep their agents running.

I think this ritual will come to an end soon because what I really want is a persistent personal computer that lives in the cloud:

An early version of this is running OpenClaw or Hermes on an always-on Mac mini. The latest version is Grok Bot, which comes closest to giving agents their own persistent cloud computer that has files, a browser, and more:

I’m sure both ChatGPT and Claude are working on a similar solution.

Your laptop’s specs will matter less when your real computer is in the cloud.

Take OpenAI’s rumored hardware device. It’s a hockey-puck-sized smart speaker that doesn’t have a screen. Soon we’ll all just be talking to agents that live in our cloud computers to get them to do work.

3. Products will be built for agents first

I’ve been trying to scratch my builder itch ever since I left my PM job. Recently, I shipped two free, open-source tools that people loved:

• No AI Slop (4,400+ GitHub stars) gives agents instructions for removing AI slop from writing.

• Human Review (570+ GitHub stars) lets people edit HTML and Markdown files and send feedback to agents, just like a Google Doc.

I think people like these products because I built them for agents first:

• Both work inside your favorite AI harness. I don’t want to use another app or website — I want tools that work with ChatGPT.

• The agent is the primary user. For example, Human Review sends feedback to the agent to make fixes.

My point is that many productivity products will look more like skills and tools for agents to use instead of websites and apps that people visit.

4. Most software will become commoditized

I laugh when I see taglines like “Anyone can build an app now!” The truth is:

Software has become easier to build but much harder to monetize.

Consider the following:

• Pieter Levels recently shared that he’s seeing declining traffic and revenue across indie SaaS products, including some of his own:

• Traditional SaaS companies are also facing pressure. Airtable sold for $1.29B after once being valued at $11B, and Canva cut its revenue forecast by 20%.

Many SaaS tools are becoming harder to justify paying for when people can just build the same thing inside ChatGPT or Claude. I recently interviewed Amol Jain (Head of Core Engineering at Replit) and asked him this spicy question:

He walked me through two examples:

1. TryNearby ($100K+ ARR) connects local businesses with creators for marketing.

2. GenAIPI ($180K in 6 weeks) helps companies assess, train, and certify employees’ AI proficiency.

Notice that neither company sells software by itself. TryNearby has a local creator network, while GenAIPI combines AI training with a trusted credential. The software makes both services easier to deliver, but it isn’t the scarce part.

Software without a real advantage, like proprietary data, distribution, or services, will become a commodity.

5. Trust will determine the winners and losers

I’ve been staying in Vancouver, and most of my friends and family here still use vanilla ChatGPT for Q&A. When I ask them why, they say: “I don’t trust AI enough to give it access to my email, calendar, and other apps.”

I believe that trust is the biggest barrier to broad agent adoption.

Cloud computers will only make the trust hurdle even higher. While testing Grok Bot, I came across this screen on its cloud computer:

If this were on my laptop, I’d probably be comfortable entering my password. But because it’s on a cloud computer, I hesitated. Before proceeding, I’d want to know what the computer remembers, who can access it, and how I can revoke access.

The winning products will make trust and transparency first-class product features.

Get a head start on this future now

To recap, I think we’ll rapidly transition from using keyboards, mice, and laptops to directing agents in the cloud using our voice. If you want to get a head start:

1. Use ChatGPT Voice to orchestrate your other threads. Talk to it and give it instructions like “Spin up a chief of staff thread to look for emails to respond to.”

2. Try Grok Bot to experience cloud agents. Connect it to a few of your favorite apps to see how a persistent cloud computer works.

3. Vibe code products for agents first. Instead of vibe coding another standalone app for humans, build a tool that agents can use. Here’s my GitHub for inspiration.

Personally, I think monitoring agents on my laptop all day is incredibly draining.

I much prefer to close my laptop lid, go outside, and talk to agents to get work done. If you haven’t tried this yet, you really should.

Today, I want to share a new episode with Nan Yu and Jacob Shumway.

Nan and Jacob built Linear Agent, which has transformed how the $1.25B company gets work done. In this episode, they walk through how they built an AI agent from the initial memo to launch, including how to give it tools to find the context it needs and how to design evals that make the agent more reliable. If you want a concrete, behind-the-scenes example of how to build a production agent end to end, this episode is a must-watch.

Watch now on YouTube, Apple, and Spotify.

We talked about:

• (00:00) What’s holding AI agents back as models improve?

• (00:51) Three things every good agent needs

• (03:00) Demo: The internal memo that started Linear Agent

• (09:50) The biggest mistake to avoid when building agents

• (11:29) Demo: From Slack to code in 6 minutes with the agent

• (17:01) When prototyping, “use the biggest model until it works”

• (24:02) Give agents tools to find context, not more context

• (34:07) The first thing to do before building an agent

I’m proud to partner with Oceans Talent

I’ve been working with Char, my assistant from Oceans Talent, for about six months. What stands out about his work is that he doesn’t just use AI tools. Instead, he actively builds and maintains AI workflows to help me run podcast post-production, draft show notes, make video clips, and more.

If you want to hire AI-fluent operators who can help you take back your time at 70% less than the cost of a US hire, book a call at oceanstalent.com/peter.

Get Started for Free

Top 10 takeaways I learned from this episode

1. The first step to building a good agent is to map the actual workflow. Identify where work starts, which systems hold context, what actions complete the job, what “done” means, and where a person needs to review. If work starts in Slack, make Slack the on-ramp instead of forcing users into a separate chatbot.

2. Give agents tools to find context instead of stuffing context into the prompt. Jacob’s advice: “Give it as little instruction as possible, to be honest. Give it the tools to load context. Don’t give it context.” Start small, then let the agent pull only the instructions and data it needs for the current task.

Read more

My last open-source skill, /no-ai-slop, went viral with 4K GitHub stars. Today, I’m introducing /human-review, another free AI skill I think you’ll love.

Let’s face it — giving feedback to AI in chat is painful. Asking AI to ‘update the button text’ or ‘edit the headline’ is annoying when I often want to edit the text myself or comment on the exact parts that I want it to fix.

/human-review opens HTML and Markdown files in a visual editor where you can edit text directly, leave comments like a Google Doc, and send feedback to AI.

Get /human-review for free on GitHub

If you find it useful, please star the repository so more people can find it. You can also upgrade to paid to get 12+ more AI skills for personal advice, agent-ready specs, viral social posts, and more.

Watch the demo below, then keep reading to learn:

1. How to install and use /human-review

2. How /human-review works behind the scenes

I’m proud to partner with Oceans Talent

I hired Char through Oceans to help with podcast production. He quickly grew beyond the original role and now creates clips, learns tools like Claude Code and Codex on his own, and delivers more every week.

Oceans Talent finds vetted operators across marketing, operations, finance, and executive support. Their hires know how to work with AI and cost one-third to one-fifth as much as comparable US talent. If you need help, book a call at oceanstalent.com/peter.

Get Started for Free

How to install and use /human-review

To install the skill, paste this into Codex, Claude Code, or your favorite AI coding tool:

After you install it, you can open the visual editor by typing:

In the visual editor, you’ll see a preview of your HTML or Markdown file on the left and a feedback panel on the right:

Now you can:

1. Edit text directly and tweak basic formatting (e.g., bold, italic).

2. Resize images by dragging their corner.

3. Give AI feedback by selecting text or a component and leaving a comment.

4. Send edits and comments in one batch by clicking the “Send X to agent” button.

Giving feedback directly inside an HTML or Markdown file instead of in a separate chat window is super useful. I use /human-review to:

1. Edit planning docs before building new features.

2. Edit and give feedback on landing pages.

3. Tweak product copy and remove AI slop from the UX.

You can also Command-click links in the visual editor to navigate between pages and leave feedback on each one.

How /human-review works behind the scenes

When you run /human-review, it starts a small local server that records your changes in the visual editor and passes them back to your AI agent.

1. Every edit, deletion, and comment is saved and grouped by page.

2. Your agent waits for feedback in 10-minute windows and receives everything immediately when you click Send.

3. Note: For HTML files, direct edits and resizes save automatically. For Markdown and localhost pages, click Send so your agent can apply them to the source.

Because the whole review loop runs locally, /human-review doesn’t require an account or API key and doesn’t upload your feedback to a separate cloud service.

Run /human-review to prevent AI from turning your project into slop

It’s easy to give AI feedback in chat and trust it to edit your documents without reviewing the actual files.

Over time, I’ve found that relying on AI to make edits can slop-ify your entire project if you’re not careful.

So the next time your agent creates an HTML or Markdown file:

1. Type “/human-review (your file)” to open it in the visual editor.

2. Make a direct edit, remove slop, or leave a comment for AI to fix.

3. Hit Send and let your agent update the source.

Get /human-review for free on GitHub and star the repo if you find it useful!

P.S. Credit to my friend Kun, whose tool Lavish inspired /human-review. I built my own version to add direct editing and Markdown support.

Today, I want to share a new episode with Karan.

Karan is the co-founder of Nous Research, the company behind Hermes, the #1 personal agent and AI app on OpenRouter. We had a great chat about why open-source AI must win and how Hermes gets better the more you use it. Karan also shared a live demo of how he used Hermes to modify his favorite childhood game.

Watch now on YouTube, Apple, and Spotify.

Karan and I talked about:

• (00:00) How Hermes differs from Claude Code and Codex

• (02:20) Why every AI model now sounds exactly the same

• (11:16) How to stop AI from saying “you’re absolutely right”

• (15:16) How Hermes creates skills without bloating its context

• (16:40) The open vs. closed-source debate on AI safety

• (23:40) Using one Kanban board for humans and AI agents

• (26:45) Demo: Rebuilding a childhood Sonic world with Hermes

• (36:07) How two Discord friends started Nous Research

• (42:40) How Hermes became its own biggest contributor

I’m proud to partner with Linear

You can now use Linear Agent to edit documents and project descriptions or set up loops to keep them updated automatically as your code changes. Linear makes agent edits transparent through author attribution and version history.

Try Linear for free

Top 10 takeaways I learned from this episode

How to make Hermes work for you

1. The “personal” in personal agent is not the model. Instead, it’s the agent’s memory of your conversations and the skills that you’ve built with it. Karan barely notices any difference in Hermes’ personality when he swaps models because his agent has accumulated so much context about him. The model is interchangeable, but the context layer is what you should cherish and preserve.

2. Customize Hermes’ soul.md to the personality that you want. After using Hermes for 6+ months now, I’ve customized its soul.md to have:

   • Core truths like be genuinely helpful and avoid sycophancy.

   • Boundaries like respecting my privacy especially in any group chats.

   • The right vibes that make it fun to talk to.

   Here’s my complete soul.md that you can copy and paste into Hermes or another AI agent to improve its personality:

Read more

Upgrade to unlock this live workshop and full access to behindthecraft.com with my best AI courses, skills, tool credits, and workshops.

Get instant access

Here’s the recording and key takeaways from our recent live workshop. Paid subscribers can find all workshop recordings here.

Start with the architecture, not the interface

Read more

Today, I want to share my six-step process to design and build Tastemaker, a website for sharing your favorite movies, TV shows, and games.

You’ll learn how to define the problem, create a design.md, prototype in Claude Design, draft an interactive HTML spec, and build the app with AI.

Watch my 25-minute tutorial to see me cover all six steps live, or keep reading below.

Timestamps:

• (00:00) My six-step AI-native design process

• (00:50) Step 1: Define the user problem before building

• (02:25) Step 2: Get inspired and create a design.md

• (05:00) Step 3: Prototype key screens in Claude Design

• (12:59) Step 4: Create an HTML spec for product, design, and tech

• (17:02) Step 5: Design the core flows, default states, and edge cases

• (20:42) Step 6: Build the app with Claude Code

• (23:07) Recap: Why planning should take at least half your time

I’m proud to partner with Replit

Replit is now available inside Claude Design, so you can design beautiful apps and send them directly to Replit to build. To try it, simply select Share in Claude Design, open the Send to… menu, and choose Replit. Get 3 months of Replit free with code PYANG26 at checkout.

Get Replit for 3 Months Free

1. Define the user problem

I’ve watched thousands of movies and TV shows and played hundreds of video games over the years. But IMDb isn’t a great interface and Letterboxd doesn’t support games.

That’s why I decided to build Tastemaker to share my favorite movies, shows, and games all on one page. To kick off the project, I asked Claude Code to research and define the user problem with this prompt:

Here’s the problem statement Claude came up with:

Note how Claude used my spec skill to find real evidence that other people share the same problem. Of course, if I wanted to turn this into a real business, I would spend more time upfront talking to users to validate the opportunity. But since I’m building this as a personal app, let’s move on to step 2.

Get my /spec skill by upgrading to a paid membership

The last skill that I open-sourced, /no-ai-slop, went viral with 3.1K GitHub stars. I’m just as proud of my /spec skill that I use throughout this tutorial. It creates an HTML spec with the product, design, and tech requirements all in one place. Get it, along with 12+ other hand-crafted skills, by upgrading to a paid membership.

2. Get inspired and create a design.md

A design.md is simply a text file with guidelines for color, typography, spacing, and more. It’s a necessary step to avoid AI generating designs that have that recognizable Claude look.

To create a design.md, I like to start by finding inspiration. Recently, I came across a beautiful app called Monogram and love how it uses color to focus people on the content instead of the UI:

I copied a few of Monogram’s screens into Claude along with this prompt to create the design.md for Tastemaker:

Claude will then create a design.md like this:

For more inspiration, you can also check out Mobbin for app screenshots and designmd.sh for design.md files for popular products.

Quick note: Getting inspiration is different from straight-up copying an app.

I feel comfortable borrowing a little from Monogram’s design because Tastemaker is in a completely different product category.

With the design.md ready, let’s move to Claude Design to prototype the product.

3. Prototype key screens in Claude Design

Claude Design is my favorite AI-native design tool because it feels more like working with a design partner than using a tool.

Tip: You don’t need Fable to prototype with Claude Design. I found Opus perfectly up to the task if you want to save tokens.

Here’s the prompt, along with the design.md, to kick off the Claude Design project:

Read more

Today, I want to share a new episode with Jason Liu.

Jason is a DevEx engineer on the Codex team. In our episode, he showed me how he runs his entire workday using Codex, from his chief of staff thread for Slack and email to setting up long-running tasks that have verifiable goals. Jason wrote the official OpenAI handbook on how to get the most out of Codex, so don’t miss this episode.

Watch now on YouTube, Apple, and Spotify.

Jason and I talked about:

• (00:00) The only job left after AI can do the work

• (02:40) Turn pinned threads into long-running workspaces

• (05:50) Build a chief of staff across Slack, email, and Linear

• (10:12) Teach Codex to write like you for every audience

• (13:33) Mine past sessions for new skills and workflows

• (20:01) How Codex revised a launch video while Jason was out

• (24:23) Build a drum practice site around how you learn

• (25:59) Goal, plan, and worklog for long-running Codex tasks

• (33:52) Why taste starts with knowing what feels wrong

I’m proud to partner with Google AI Studio

Google AI Studio was my go-to prototyping tool when I was a product manager. You can use it to generate three versions of your product, add visuals with Google’s Nano Banana model, and publish them to a custom URL for customer feedback. It’s an easy, free way to move from prompt to something your team or customers can try.

Try Google AI Studio for Free

Top 10 takeaways I learned from this episode

Build a Codex system that improves itself

1. Build a chief of staff Codex thread that manages your email, Slack, and more for you. At 9 am, 1 pm, and 5 pm, Jason’s chief of staff thread reviews his Slack, emails, Twitter DMs (through browser use), and Linear before surfacing priorities with links and draft replies. Here’s a prompt you can modify to set this up:

2. Ask Codex to look at your past sessions to build new skills and workflows. Codex has memory of all your past conversations, so the best way to save time is to ask it to find workflows from your manual iterations that it can build new skills to automate. Here’s a prompt you can paste to do this right now:

Read more

Today, I’m open sourcing my /no-ai-slop skill that removes 20+ patterns of AI slop from any piece of writing.

I use this skill to edit everything from emails to docs, and I think you’ll love it too.

Get the skill for free on Github

If you find it useful, please consider starring the repository so more people can find it and upgrading to paid to get 12+ other quality AI skills I’ve made for personal advisor, agent-ready specs, viral social posts, and more. Keep reading for:

1. How to install and use the /no-ai-slop skill

2. My 25/50/25 rule for editing with AI

3. Why the dark side of slop is always calling

I’m proud to partner with Linear

I’m using Linear to manage my entire podcast production process with my video editors and it’s been great. With Linear MCP, I can create, edit, and track issues directly from Codex and Claude Code.

Try Linear for free

How to install and use the /no-ai-slop skill

To install the skill, just paste this in Codex, Claude Code, or your favorite AI harness:

Then you can use it in three ways:

1. “/no-ai-slop (text)” to edit your writing. The skill will remove all the AI slop patterns and list what it removed.

2. “/no-ai-slop is this slop? (text)” to detect AI writing. The skill will list all the slop patterns that it found.

3. “draft an AI slop post (text)” for some fun. Use it to generate the most cringe slop as satire like my LinkedIn comment below:

Jokes aside, even the best AI models (e.g., Fable 5 and GPT-5.6) still use most of the slop patterns below that /no-ai-slop will remove:

1. Binary contrasts. “It’s not X. It’s Y.”

2. Throat-clearing openers. “Here’s the thing,” “Let me be clear”

3. Faux-insight setups. “What nobody tells you,” “The part everyone misses”

4. Colon reveals. “The best part: it learns.”

5. Dramatic fragments. “That’s it. That’s the whole thing.”

6. Superficial analysis. “highlighting the team’s commitment to innovation”

7. Importance puffery. “marks a pivotal moment,” “a testament to”

8. Weasel attribution. “experts agree,” “studies show”

9. Synonym cycling. “The agent handles your email. The assistant drafts replies.”

10. Fake-profound endings. “The future isn’t coming. It’s already here.”

Here’s an example of using the skill to detect AI slop on a recent tweet I came across:

My 25/50/25 rule for editing with AI

Substack just shipped AI slop detection to protect readers. They featured my video as an example of how to use AI to edit while still sweating the details.

Basically, I work with AI in the middle 50% of my writing and editing process:

1. First 25%: I start by voice dictating or writing a rough draft without AI.

2. Middle 50%: I use my /newsletter and /no-ai-slop skills to work with AI to clean up the draft. AI helps me find grammar mistakes, improve clarity, and more.

3. Last 25%: I manually go through the piece line by line to make edits without AI.

I’d like to think that the difference between a creator who pumps out slop and a creator who cares about quality is giving a shit about the first and last 25%.

This applies to building with AI too. As Kun (AI builder) shared in our interview, AI handles the middle 50% where most of the code gets written. But you have to care enough about the first 25% (using your taste to shape the requirements) and the last 25% (validating the work) to make a product great.

Why the dark side of slop is always calling

The dark side of just letting AI draft and edit everything is always calling:

1. It’s easier to just give AI feedback to edit than to do it manually. Products like Codex and Claude Code seem to actively discourage you from making manual edits to files.

2. It’s easier to just be lazy and not write at all if AI is not available. For example, I used to be productive on long flights. But lately, I’ve become too lazy to write or edit anything without having AI available.

Perhaps most depressing of all, pumping out slop can drive real results.

Without naming names, I’ve seen some creators publish 10 AI slop posts per hour, every day. Some of those posts go viral and bring them more followers and revenue. But I like to remind myself that:

People who know AI can see the slop coming from a mile away.

A post can perform while still costing you trust with the people whose judgment matters to you. Consider these two examples:

1. Brian Chesky’s account got hacked and posted an obviously AI generated thread. Readers spotted it as AI before anyone confirmed the hack.

   

2. My friend, Alex Cohen, wrote a letter to his employees highlighting that pumping out slop puts an unfair burden on the reader.

Don’t let AI make you dumber and lazier

I’m not going back to editing everything manually, but I also don’t want to lose my ability to think or craft a sentence when there’s no Wi-Fi available.

So I’m going to hold the line at my 25/50/25 writing process. Trust is the only durable asset we all have, and slop hurts trust with the people who matter.

Get my free /no-ai-slop skill on Github and let me know what you think!

Get instant access

Here’s the recording and key takeaways from our recent live workshop. Paid subscribers can find all workshop recordings here.

How Julie’s AI-native design workflow works

Read more

Today, I want to share a new episode with Thariq Shihipar.

Thariq works on the Claude Code team, and I’ve wanted to see how he builds for a long time. In our episode, he showed how to use /goal to keep Claude working, how he plans with Claude to remove unknowns before building, and how he runs a team of agents in Slack. He also shared why his team cut Claude Code’s system prompt by 80% and how he creates polished launch videos with Claude.

Watch now on YouTube, Apple, and Spotify.

Thariq and I talked about:

• (00:00) Goals, loops, and workflows to make Claude work longer

• (04:48) Live demo: Creating a polished video with Claude

• (08:17) Plan by removing unknowns before you build

• (14:32) How Anthropic uses HTML artifacts for planning

• (18:35) Running parallel Claude tasks through Slack

• (27:34) Using separate creator and verifier sub-agents

• (31:08) How Thariq stays focused while Claude works in parallel

• (34:07) Why Anthropic cut Claude Code’s system prompt by 80%

• (37:01) How to become more technical when AI writes the code

I’m proud to partner with Wispr Flow

When I’m building with Claude Code, I use Wispr Flow to dictate all my prompts instead of typing. It’s more accurate than other AI voice dictation tools and saves me hours every week. Try it out for free with my exclusive link below.

Try Wispr Flow for Free

Top 10 takeaways I learned from this episode

1. When a better AI model is released, the first thing you should try is to remove instructions. Thariq and the Claude Code team cut Claude Code’s system prompt by 80% because hard rules and examples were constraining newer models. When a new model is released, audit your CLAUDE.md and skills to remove deterministic instructions.

2. Plan with Claude to discover and remove unknowns. Before building his one-shot video workflow, Thariq asked Claude to explain how voice transcription tools like Whisper could fail. This helped him clarify the plan before getting Claude to create the video. Perhaps the most important learning here is that he actually reads the plan manually instead of just trusting AI.

3. Use /goal only if you have a verifiable finish line. I asked Claude to “/goal Build me a great game” and the results were…not great. Thariq shared how to use /goal to create polished videos with Claude by providing a verifiable end state (e.g., a rendered video). Here’s the exact prompt he used to create the video in one shot:

Read more

Get instant access

Dear subscribers,

I now use ChatGPT Work and GPT-5.6 for almost everything I do on my computer, and I’m going to show you my full setup in this video and written guide.

You’ll learn how to choose the right GPT-5.6 model, connect your apps, manage email and calendar, automate recurring work, build Sites, and more.

Watch my tutorial now for the full 28 min walkthrough, or keep reading below.

Timestamps:

• (00:00) How I use ChatGPT Work for almost everything

• (01:04) Chat vs. Work vs. Codex: What’s the difference?

• (02:22) Which GPT-5.6 model and effort to use

• (03:18) Personalizing ChatGPT Work’s custom instructions

• (06:23) Connect your apps and organize your Work threads

• (09:30) Let ChatGPT manage email and draft replies in your voice

• (13:47) Schedule events and prep for meetings automatically

• (17:19) Automate recurring work with scheduled tasks

• (19:23) Continue your Work tasks from your phone

• (21:01) Publish travel guides, games, and product plans as Sites

I’m proud to partner with Wispr Flow

I use Wispr Flow to dictate first drafts much faster than I can type. It helps me get my ideas down while they’re still fresh, so I can focus on editing. I use it all the time in ChatGPT Work and Codex, and it’s been a huge time saver. Try it for free with my link.

Try WisprFlow for Free

Chat vs. Work vs. Codex: What’s the difference?

Here’s the simplest way to think about the three modes:

1. Chat is the regular ChatGPT we’re all familiar with.

2. Work lets ChatGPT take action in your favorite apps.

3. Codex adds features designed for coding and developers.

Honestly, I’m not a fan of Work and Codex existing as separate modes. As far as I can tell, there are only small differences in the UI and how responses are presented. For example, Codex adds a Pull Request tab that developers can use to review each other’s code.

So I don’t think it makes much difference whether you use Work or Codex. My recommendation is simple:

Use the ChatGPT desktop app which is far more capable than ChatGPT for the web.

Which GPT-5.6 model and effort should you use?

GPT-5.6 comes in three variants (Sol, Terra, and Luna), each with five effort levels from Low to Ultra. So which model and effort should you default to?

Here’s a chart that shows intelligence vs. output tokens for the various options:

Based on this chart and my experience, I recommend that you:

Use GPT-5.6 Sol (medium) for the best balance of intelligence and token use.

I sometimes use Sol (high) for planning, but overall I haven’t noticed much of a difference between that and medium. I almost never use Terra or Luna.

Personalize your ChatGPT Custom Instructions

On a fresh install, I go to Settings → Personalization and set my Custom Instructions to:

1. Teach me technical concepts without using jargon.

2. Challenge my thinking instead of agreeing with every prompt.

3. Remove AI slop phrases and patterns.

Here are my exact custom instructions if you want to copy my setup:

Install plugins to let ChatGPT use your favorite apps

Plugins connect ChatGPT Work and Codex to the apps you use every day.

The screenshot above shows some of the plugins I have installed. I think the bare minimum set is:

• Gmail

• Google Calendar

• Documents, Presentations, Spreadsheets, and Google Drive

Other useful plugins include the PDF plugin for editing PDFs, Slack for messages, Notion for wiki articles, Figma for design, and Granola for meeting transcripts.

Install plugins for any app that you want ChatGPT to use. This is the key setup step that makes ChatGPT Work and Codex your everything app.

Use long-running threads for each type of work

As you use ChatGPT Work more, you may get overwhelmed by all your active projects and threads. To avoid this:

Keep one long-running thread for each type of work.

For example, in my Personal OS project, my threads include:

1. Newsletter to edit all newsletter posts.

2. Social for drafting and making posts for social networks.

3. Podcast to prep and do post-production for interviews.

ChatGPT Work and Codex are very good at compacting long conversations, so you can keep chatting in the same thread without worrying about GPT getting confused.

I also pin my most-used threads, such as my Advisor thread for strategy and business advice and my Chief of Staff thread for email, calendar, and more.

Set up a chief of staff thread for email and calendar

These days, I manage my email and calendar almost entirely through ChatGPT Work and Codex instead of opening Gmail or Google Calendar. Let me walk through how I:

1. Find forgotten follow-ups and subscriptions to cancel

2. Draft email replies that sound like you

3. Manage your calendar and avoid meeting conflicts

4. Get briefed on upcoming meetings

1. Find forgotten follow-ups and subscriptions to cancel

Read more

Get instant access

Dear subscribers,

Today, I want to share a new episode with Jared Zoneraich.

Jared is a Builder in Residence at Cognition, the team behind Devin. He showed me why teams often overbuild their first agents, and how simple prompts, better tools, and self-testing work together instead. Then he demoed one manager agent breaking down a task, directing 10 cloud agents, and deciding what a human needs to review.

Watch now on YouTube, Apple, and Spotify.

Jared and I talked about:

• (00:00) Why teams overbuild their first AI agent

• (02:50) Let the model cook before adding rules

• (03:16) Why tools matter more than better prompts

• (10:51) Why you should build evals after shipping

• (16:06) What cloud agents can do that local agents can’t

• (20:30) How to get one agent to manage other agents

• (25:39) Demo: One agent launches a team of 10

• (28:03) How agents prove their work before you review it

• (33:18) The best tasks for agent teams to work on

I’m proud to partner with Riverside

I produce a podcast every week and it used to take me at least 10 hours to record, edit, cut clips, write posts, and publish everywhere. Riverside has cut that time by half.

In addition to recording at 4K local with clean audio, I can also now edit using the transcript, add captions, make clips, and create and publish social posts all from one place. Riverside is trusted by teams like Spotify, Microsoft, and Vercel and you can try it for one month free at creators.riverside.com/PeterYang with code PeterYang.

Try Riverside for Free

Top 10 takeaways I learned from this episode

How to build great AI agents

1. Start by expressing your intent to the model in a simple prompt. “Let the model cook” by expressing your intent and constraints clearly instead of forcing it to follow deterministic instructions.

2. Give the model relevant tools to pull context. “Tool engineering” is giving the model ways to search your code, read logs, run tests, open the app, and inspect what it changed.

Read more

OpenAI just made GPT 5.6 available to everyone and I’m here to answer your most burning question:

Is GPT 5.6 better than Claude Fable 5?

The answer is, it depends. 😅

Watch my tutorial to see me test both models head-to-head for building a travel website, 3D Star Fox game, mobile app, and more.

Timestamps:

1. (00:00) GPT-5.6 vs Claude Fable 5 compared

2. (01:45) Build an interactive travel website

3. (05:44) Create a 3D Star Fox-style game

4. (09:39) Edit and publish short videos with browser use

5. (15:10) Plan and build a mobile app feature

6. (19:34) Get personalized life and business advice

7. (22:17) Improve and clean up a personal AI OS

8. (24:37) My takeaway on each model’s strengths and weaknesses

I’m proud to partner with Riverside

Right now, most creators spend 20% of their time creating, and 80% on editing and distribution. Riverside just launched 2.0 to make it easy for me and other creators to turn a single recording into a fully produced content engine across all my channels.

After recording, Riverside uses a team of agents to create clips, schedule social posts, and more. Riverside is trusted by teams like Spotify, Microsoft, and Vercel, and you can try it for one month free at creators.riverside.com/PeterYang with code PeterYang.

Try Riverside for 1 Month Free

GPT 5.6 vs. Fable benchmark and cost comparison

GPT 5.6 comes in three models: Sol, Terra, and Luna.

Sol is the flagship model and based on TerminalBench it actually performs slightly better than Claude Fable 5 at coding:

But I think OpenAI’s biggest advantage is model efficiency and cost. GPT 5.6 Sol is 50% cheaper than Claude Fable 5 at API prices:

That’s a big deal because enterprises are increasingly looking for the most cost-effective model. However, benchmarks only tell half the story, let me show you how both models perform on the 6 use cases below:

1. Build an interactive travel website

GPT has historically lagged behind Claude on front-end design so I wanted to see if GPT 5.6 has closed that gap. I’m planning a trip to Japan in December so I gave both models my Google Doc itinerary and the prompt below:

Tip: Ask AI to “add a 3D WebGL element to the hero banner” if you want to create websites with beautiful animated 3D graphics

I ran the same prompt through GPT 5.5, GPT 5.6, and Fable:

1. GPT 5.5 created a functional website but botched the 3D hero image.

2. GPT 5.6 created an animated torii gate and found more relevant photos.

3. Fable 5 created beautiful animated snowflakes in a winter scene.

Watch my video to see how all 3 websites look.

Verdict: Tied because GPT 5.6 has really closed the gap with Fable on frontend design.

P.S. We’ve started hosting regular live workshops in my Slack community for AI builders. Upcoming workshops include:

1. July 17 - AI native design (design.md, components, etc) with Julie Xu

2. July 24 - Building production apps with Colin Matthews

3. August - Deep dive on AI evals with Hamel Husain

All workshop recordings will be shared here for paid subscribers.

Subscribe for Full Access

2. Create a 3D Star Fox style game

Growing up, Star Fox was one of my favorite games on the Super Nintendo. So naturally, I had to ask both GPT 5.6 and Claude Fable 5 to build a fully playable Star Fox level complete with enemies, power-ups, and even a boss battle.

Both produced incredibly impressive results:

Here’s the exact prompt that I used to build these games in one shot:

Read more

Get instant access

Dear subscribers,

Today, I want to share a new episode with Rohan Varma.

As a product manager on OpenAI Codex, Rohan has a firsthand look at how one of the most AI-native teams in the world works. He showed me how he uses Codex to automate work from Slack replies, prototype designs with image generation, replace stale PRDs with live sites, and more. Watch this episode if you’re tired of all the nonsense around “AI PM” and want to see how a real one actually operates.

Watch now on YouTube, Apple, and Spotify.

Rohan and I talked about:

• (00:00) Codex can do so much more than you think

• (01:53) Use Codex to pull context from Slack, docs, Linear, and email

• (03:28) How Codex flips planning from docs to demos

• (13:14) Demo: Do the task once with Codex, then automate it

• (14:30) Demo: Trigger Codex automations from Slack replies

• (17:26) Demo: Use Image Gen to explore design variants fast

• (19:58) Demo: Turn any Codex thread into a reusable skill

• (29:55) Demo: Turn project context into a shareable Site

• (33:33) Use one Codex thread to manage other Codex threads

• (37:18) Aim to be 10x more ambitious than feels reasonable

I’m proud to partner with Oceans Talent

I’ve been working with my Oceans Talent Executive Assistant for 3 months now and he’s helped me streamline my entire content pipeline from research to podcast post-production. He builds his own AI skills to save both of us hours every week. He’s more than an assistant and runs the whole system.

Oceans Talent places vetted, AI-fluent Executive Assistants who go beyond inbox and calendar, with backgrounds in marketing, operations, and finance so they can run your projects and own outcomes. Book a call today to get your first month free.

Get Started for Free

Top 10 takeaways I learned from this episode

Set up Codex as your work layer

1. Connect Codex to the tools where your work already happens. Rohan connects Codex to Slack, Notion, Linear, Gmail, Drive, and other tools. The magic starts when it can pull context from multiple connected tools and navigate to webpages that it has permission to access.

2. Use Codex to get up to speed on any project fast. Rohan said he can get added to a project last minute and quickly catch up by asking Codex to gather the Slack history, docs, Linear issues, email, and Drive context he is allowed to access.

3. Codex handles the manual work so PMs can spend more time with users. Since Codex can handle the manual PM work around information synthesis and follow-ups, Rohan has more time now to spend meeting enterprise customers and teams.

Automate the repetitive work

Read more

Get instant access

Dear subscribers,

Today, I want to share 18 hot takes on where I think the AI market is headed.

AI is in a weird place right now. The government is restricting access to frontier models, enterprises are becoming conscious of token costs, and everyone’s trying to rebuild their product for agents first instead of humans.

I’ve interviewed dozens of AI leaders and spent far too much time following these topics on X/Twitter. Here are 18 hot takes on where I think AI is headed next:

1. The frontier-only AI stack is collapsing

2. The AI super app era is here

3. Traditional software risks becoming a dumb pipe for agents

4. Cloud agents and collaboration are the next wave

I’m proud to partner with Oceans Talent

My Oceans Talent Executive Assistant started by doing simple admin tasks but has now built his own AI skills to turn podcast episodes into clips, social posts, and more. Sourcing someone this capable on my own would have taken months.

Oceans Talent is execution capacity you can hire this week. They provide vetted, AI-fluent executive assistants with backgrounds in marketing, operations, and finance to run your projects and own the work. Book a call today to get your first month free.

Get Started for Free

The frontier-only AI stack is collapsing

1. Tokenmaxxing at frontier API prices makes no sense. Uber burned through its entire 2026 AI budget in 4 months, Microsoft moved engineers off Claude Code due to cost, and companies are realizing that running everything on frontier models can get expensive fast. Tokenmaxxing makes sense when you’re on a subsidized $200/month plan but is unsustainable at API rates.

2. Companies will rely on a portfolio of models. Coinbase recently cut its AI spend nearly in half by switching engineers to Chinese open-source models like GLM and Kimi. Airbnb and Pinterest have done the same with Alibaba’s Qwen models. I believe that this will be the default path forward — using frontier for high-stakes work and cheaper models for everything else.

   

3. China’s open-source strategy is working. Chinese models are taking market share from frontier models at US companies. China is also building the full AI stack — from energy (e.g., solar, nuclear) to data centers to domestic chips. The Chinese government is planning a $295B investment in AI data centers with at least 80% of the chips built domestically.

4. Frontier labs are in a catch-22 situation. If they release great open-source models, they might undercut their own frontier API revenue. If they gate the best models behind a trusted list, companies will just lean on open alternatives more. The last major US open-source model was OpenAI’s gpt-oss series back in August 2025, which already feels like decades ago in the AI space.

   

5. The US needs to think about its AI strategy holistically. I believe that restricting access to frontier models will only hurt American innovation. Banning US companies from Chinese models won’t work either. Just look at how China took over the global electric vehicle market. To compete, we need the best closed and open models while scaling our energy grid and data centers much faster.

The AI super app era is here

6. Anthropic and OpenAI want to transform all knowledge work. Both labs want to remake how white-collar work gets done with agents. What started with coding has expanded to design, finance, legal, science, and more. Knowledge workers are now adopting Codex three times faster than developers.

7. The race to build the best AI super app is on. Whoever wins it will take massive market share away from pre-AI knowledge work tools. The main players are Codex, Claude, Cursor, and a few others. With browser and computer use, I can already get Codex to handle almost everything I want to do on my computer.

8. OpenAI is well positioned to win the super app race. ChatGPT has 800M+ weekly active users and Codex is arguably the best AI super app right now. OpenAI has a clear opportunity to merge ChatGPT and Codex into one app and push agents into the mainstream. I think this is imminent.

   

9. Anthropic leads enterprise but has a fragmented product suite. Long term, I don’t think it makes sense for Claude, Cowork, Claude Code, and Claude Design to all exist as separate products. Working with Claude needs to feel more like working with a single capable coworker instead of a product suite. Claude Tag, Anthropic’s new agent for Slack, is a step in the right direction.

Software risks becoming a dumb pipe for agents

10. Agents are becoming the default user. Most products are still designed for humans clicking buttons, but agents will very soon read, write, buy, and do a lot more on our behalf. Cloudflare’s data shows that bot traffic to websites already exceeds human traffic. This matches my personal experience — I rely on agents like Hermes and Codex to browse websites and steer traditional apps.

    

11. If your product isn’t built for agents, it might as well not exist. Without clear APIs, documentation, and CLIs or MCPs, agents simply won’t discover your product, and neither will the humans who now lean on agents to find things. That’s why the best AI builders I’ve interviewed are all building for agents first.

12. Existing knowledge work products risk becoming dumb pipes. There’s a dark side to building everything for agents first. Take Google Workspace, for example. If I can just use Codex to edit Google Docs for me, I might never touch the Google Docs interface. Today’s apps risk becoming tomorrow’s dumb pipes for agents, the same way cable became the dumb pipe for internet apps.

13. Adding an AI chatbot to your product isn’t going to save you. A chatbot trapped inside a single app, with only that app’s context, can’t compete with an agent like Hermes or Codex that carries my personal context and gets work done across many apps. For example, I get a daily morning brief from Hermes that pulls info from my calendar, unread emails, and Slack.

14. Pure-play SaaS is losing to services with SaaS bolted on. Companies don’t want another SaaS tool to onboard to and maintain. Instead, they want an outcome delivered to them. Pure-play SaaS will struggle to compete with AI skills that users and teams can build and personalize inside their favorite AI super app.

    

Cloud agents and collaboration are the next wave

15. Startups can compete with AI labs by going deep and using a portfolio of models. For example, coding startups like Devin and Factory are doing well because their harnesses route across multiple models to help customers save on cost without giving up performance. They’re also focused on a few specific use cases instead of trying to be everything to everyone.

16. Always-on cloud agents will be accessible from any device. Running agents with your laptop half-open will feel silly soon. Instead, many builders are already launching cloud agents using Cursor, Devin, and others. Codex and Claude Code are no doubt working on the same experience, and this will be the default soon.

17. Collaboration between agents and teams has yet to be figured out. The default agent harness with a list of threads on the left is still very much a single-player experience. The multiplayer experience is still being defined — whether that’s Claude Tag in Slack, agents in Linear, or something else.

18. “The future isn’t 20 terminal tabs. It’s better loops.” I watched Peter Steinberger share this in his keynote at yesterday’s AI Engineer conference. He no longer watches the agents code. Instead, he defines the requirements, and let agents plan, execute, and validate by themselves. I’m a bit wary of loops since they can lead to AI slop, but I’m sure better planning and models will make them more mainstream.

The AI market is about to get a lot more interesting

Here are my 18 takes summarized in four bullets:

1. The frontier-only AI stack is collapsing. Companies will use a portfolio of models to save costs, with everyday tasks defaulting to low-cost open models from China.

2. The AI super app era is here. Codex, Claude, Cursor, and others are competing to disrupt knowledge work by using agents for everything you can do on a computer.

3. Traditional software risks becoming a dumb pipe. Agents will become the default user. They’ll use APIs and browser use to access your website and app without humans ever seeing them.

4. Cloud agents and collaboration are the next wave. The future is agents living in the cloud, accessible from any device, and collaborating closely with both you and your team. This will happen very soon.

Let me know what you think about these takes in the comments!

Today, I want to share a new episode with Jess Yan.

As product lead at Anthropic, Jess has a first-hand look at how Anthropic teams are moving from prompting AI to building long-running agents that work overnight. In our episode, Jess showed me how to build a Claude analytics agent from scratch, then covered how Anthropic PMs use agents internally to understand the codebase, monitor customer feedback, and make better product decisions.

Watch now on YouTube, Apple, and Spotify.

Jess and I talked about:

• (00:00) AI agents that work while you sleep

• (03:46) Claude Managed Agents help for long-running tasks

• (05:22) Demo: Building a Claude analyst agent from scratch

• (09:28) Traces: How to see where an agent got stuck

• (15:11) Evals: How agents grade their own work

• (20:20) 5 ways in which Anthropic PMs use agents internally

• (28:06) Processing a 4,000-company waitlist with agents

• (32:50) To build your first agent, focus on helping one person

• (36:54) What changes when agents can work overnight

Evan helps you be the most prepared person in a meeting

Evan is built to prepare you for high-stakes meetings. It pulls context from your apps, does CIA-level research on every attendee, and hands you a presidential-style briefing before you walk in. Built for founders, sellers, VCs, and anyone whose meetings are consequential, because showing up unprepared isn’t just awkward, it’s costly.

Get Your First Briefing Now

Top 10 takeaways I learned from this episode

Jess and I covered two topics in the interview:

1. How Anthropic PMs use agents internally

2. How to build long-running Claude agents

How Anthropic PMs use agents internally

Anthropic PMs use agents internally for at least five PM workflows:

1. Understanding the codebase. Instead of asking engineers what shipped or how something works, Jess can inspect pull requests, deployments, and product behavior directly. “Access to our codebase has been the biggest unlock for me.”

2. Synthesizing user feedback from multiple channels. Jess has agents monitor customer Slack channels to get more signal before deciding what to build next.

Read more

Today I’m excited to launch my beginner-friendly Hermes course with 17 opinionated lessons to set up my favorite AI chief of staff in a single afternoon.

This course covers everything from installation to integrations (e.g., email, calendar, Google Workspace) to cron jobs (e.g., morning briefing, health check).

Watch my free 45-minute walkthrough on YouTube. If you want the written version with all 17 lessons and copy-paste skills, consider upgrading to paid at behindthecraft.com.

What the course covers:

• (00:00) Why Hermes is the best 24/7 AI chief of staff

• (01:16) Where and how to install Hermes safely

• (02:55) The best AI model to pick for Hermes

• (05:04) Connect to Telegram for mobile chats

• (11:08) How Hermes uses tools, memory, and skills

• (12:39) How to personalize Hermes by editing its user and soul files

• (15:49) Two best ways to troubleshoot Hermes

• (18:15) Set up Hermes to respond via voice

• (22:47) Connect Gmail, Calendar, Docs, Drive, and Sheets step by step

• (38:25) Create routines like morning briefing, health check, and more

Why Hermes and what’s inside the course?

Hermes is, in my opinion, the best AI chief of staff available today. It’s reliable, free to use (you just need a $20 ChatGPT subscription), and very powerful.

That’s why I spent 20+ hours writing 17 opinionated lessons covering three tracks:

1. Quickstart: Get Hermes running in less than 30 minutes.

2. Integrations: Let Hermes work in your favorite apps (e.g., email, calendar, docs).

3. Cron jobs: Set up morning briefings, health checks, business reports, and more.

Upgrade to Get My Full Hermes Course

Here’s what each track covers.

Quickstart: Get Hermes running in less than 30 minutes

Learn how to install Hermes safely, connect it to your phone via Telegram, and personalize it to be yours by editing its user and soul files.

This track covers:

1. Where to install: Set up Hermes on your laptop, Mac Mini, or virtual server.

2. Install Hermes: Download the desktop app and pick your AI model.

3. Connect Telegram: Create your bot and chat with Hermes from your phone.

4. How Hermes works: A beginner-friendly guide to memory, tools, and more.

5. Personalize Hermes: Edit its user and soul files to make it genuinely useful.

6. Troubleshooting: My two favorite ways to fix Hermes if you get stuck.

Integrations: Connect the apps you already use

Integrations are what make Hermes more useful than a chatbot. Start with voice replies and Google Workspace, which I think are must-haves. Then consider linking Granola for meetings, GitHub for backups, YouTube for research, and more.

This track covers:

7. Voice replies: Talk to Hermes while on the go.

8. Google Workspace: Connect Gmail, Calendar, Docs, Sheets, and Drive.

9. WhatsApp: Use Hermes from WhatsApp with a second phone number.

10. Slack: Add Hermes to Slack for threaded discussions.

11. Granola: Let Hermes analyze your meeting notes.

12. GitHub: Back up Hermes and let it work with your repos.

13. YouTube: Let Hermes summarize videos and research for you.

Cron jobs: make Hermes proactive

Last but not least, cron jobs (or scheduled routines) let Hermes do work for you proactively. I cover how to set up my favorite cron jobs including:

14. Weekend planner: Find fun, time-sensitive family activities for the weekend.

15. Morning briefing: A daily check of your email, calendar, and more.

16. Business review: A weekly insights email with your financials and more.

17. Health check: A weekly review with your weight, workouts, and more.

My goal is to create the most practical AI membership online

With today’s update, paid subscribers can now access all of the following on the Behind the Craft membership portal:

1. 12+ skills and 20+ prompts handcrafted for personal, product, and creator work.

2. 17 lessons on setting up your AI chief of staff with Hermes.

3. Full Codex and Claude Code courses coming later this year.

We’ve also started hosting monthly expert-led live AI workshops in our AI Slack community, starting with my walkthrough of my Codex setup last Friday.

Again, I’ve made the full 45-minute walkthrough free on YouTube to make it easy for you to set up Hermes. If you’d like the written course, my AI skills, live workshops, and upcoming Codex and Claude Code courses, consider upgrading to paid.

Thanks!
Peter