Today’s Episode

ChatGPT Apps might be the next billion-dollar opportunity.

Or they might be another ChatGPT feature that gets abandoned in 6 months.

I genuinely don’t know yet.

But when people say “this could be the new App Store,” my ears perk up. I spent four years building an iOS app in the early days of the App Store. The distribution was incredible. We grew fast purely because of where we were.

So when OpenAI announced the ChatGPT App Store, I needed to understand it.

I brought in Colin Matthews to break it down. Colin is one of my go-to sources for technical product topics. Our AI prototyping collaborations have been some of your favorite episodes.

Today, we’re exploring ChatGPT Apps and what they mean for you as a product builder.

----

Check out the conversation on Apple, Spotify and YouTube.

Brought to you by:

* Maven: Get $500 off with my code on Coil Build ChatGPT Apps course

* Vanta: Automate compliance, Get $1,000 with my link

* Land PM job: 12-week experience to master getting a PM job

* Mobbin: Discover real-world design inspiration

* NayaOne: Airgapped cloud-agnostic sandbox

----

If you want access to my AI tool stack - Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin - for free, grab Aakash’s bundle.

Are you searching for a PM job? Join me + 29 others for an intensive 12-week experience to master getting a PM job. Only 23 seats left.

----

Key Takeaways:

1. ChatGPT apps = MCP + widgets - The Model Context Protocol (invented by Anthropic) lets AI agents call external tools. OpenAI added UI widgets on top to create embedded app experiences directly in chat.

2. 900M weekly active users = massive distribution opportunity - This is the new SEO. Early data shows 26% higher conversion from AI traffic vs traditional search. Every enterprise will eventually build here.

3. You're building for multiple platforms - MCP works across ChatGPT, Claude (coming soon), Cursor, and other AI tools. Build once, distribute everywhere. Gemini doesn't support it yet.

4. Apps get called based on tool descriptions - Your metadata matters. Like SEO but for LLMs. Run evals to test if correct prompts trigger your tools. Iterate on descriptions to improve discovery.

5. Three eval categories: direct, indirect, negative - Direct: user names your app. Indirect: user describes outcome. Negative: irrelevant request shouldn't trigger your tool. Test all three systematically.

6. PMs should prototype but engineers ship production - Use tools like Chippy to prototype quickly and test concepts. Show stakeholders real interactions. Engineering team builds the production version.

7. Enterprise-first, solo builders second - Large companies (Target, Uber, Canva) are early adopters chasing distribution. But huge opportunity for indie builders once public marketplace launches.

8. Best opportunities: embedded collaboration tools - Spreadsheets, task lists, whiteboards where ChatGPT can partner with you. Not just search results—actual interactive experiences.

9. Error analysis on observability logs is critical - Track what prompts triggered which tools with what parameters. Look for mismatches between expected and actual behavior. Iterate tool descriptions.

10. Marketplace launching by end of 2024/early 2025 - Currently only launch partners can publish. Public marketplace coming soon means anyone can ship apps and reach ChatGPT's massive user base.

----

Where to Find Colin

* LinkedIn

* Newsletter

Related Content

Newsletters:

* AI Prototyping Tutorial

* How to Build AI Products

* AI Product Strategy

* Complete Course: AI Product Management

Podcasts:

* AI Prototyping for PMs

* How to Become an AI PM

* Everything You Need to Know About AI

----

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

If you want to advertise, email productgrowthppp at gmail.

Today’s Episode

Everyone’s demoing AI features. Few are shipping them to production reliably.

The gap? Evals.

Not the theoretical kind. The real-world kind that catches bugs before users do.

Hamel Husain and Shreya Shankar train people at OpenAI, Anthropic, Google, and Meta on how to build AI products that actually work. Their Maven course is the top-grossing course on the platform.

Today, they’re walking you through their complete eval process.

----

Brought to you by:

* The AI Evals Course for PMs & Engineers: You get $800 with this link

* Vanta: Automate compliance, Get $1,000 with my link

* Jira Product Discovery: Plan with purpose, ship with confidence

* Land PM job: 12-week experience to master getting a PM job

* Pendo: the #1 Software Experience Management Platform

----

If you want access to my AI tool stack - Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin - for free, grab Aakash’s bundle.

Are you searching for a PM job? Join me + 29 others for an intensive 12-week experience to master getting a PM job. Only 23 seats left.

----

Key Takeaways:

1. AI evals are the #1 most important new skill for PMs in 2025 - Even Claude Code teams do evals upstream. For custom applications, systematic evaluation is non-negotiable. Dog fooding alone isn't enough at scale.

2. Error analysis is the secret weapon most teams skip - Looking at 100 traces teaches you more than any generic metric. Hamel: "If you try to use helpfulness scores, the LLM won't catch the real product issues."

3. Use observability tools but don't depend on them completely - Brain Trust, LangSmith, Arise all work. But Shreya and Hamel teach students to vibe code their own trace viewers. Sometimes CSV files are enough to start.

4. Never use agreement as your eval metric - It's a trap. A judge that always says "pass" can have 90% accuracy if failures are rare. Use TPR (true positive rate) and TNR (true negative rate) instead.

5. Open coding then axial coding reveals patterns - Write notes on 100 traces without root cause analysis. Then categorize into 5-6 actionable themes. Use LLMs to help but refine manually.

6. Product managers must do the error analysis themselves - Don't outsource to developers. Engineers lack domain context. Hamel: "It's almost a tragedy to separate the prompt from the product manager because it's English."

7. Real traces reveal what demos hide - Chat GPT said the assistant was correct but missed: wrong bathroom configuration, markdown in SMS, double-booked tours, ignored handoff requests.

8. Binary scores beat 1-5 scales for LLM judges - Easier to validate alignment. Business decisions are binary anyway. LLMs struggle with nuanced numerical scoring.

9. Code-based evals for formatting, LLM judges for subjective calls - Markdown in text messages? Write a simple assertion. Human handoff quality? Need an LLM judge with proper rubric.

10. Start with traces even before launch - Dog food your own app. Recruit friends as beta testers. Generate synthetic inputs only as last resort. Error analysis works best with real user behavior.-

---

Where To Find Them

* LinkedIn:

* Hamel: Hamel’s LinkedIn

* Shreya: Shreya’s LinkedIn

* AI Evals Course: World’s best AI Evals Course (You get $800 off with this link)

Related Content

Newsletters:

* AI Evals

* AI PM Observability

* AI Testing

* LLM Judge

* AI Prototype to Production

* AI Product Strategy

* How to Build AI Products

* Prompt Engineering

Podcasts:

* AI Evals: Everything You Need to Know to Start

* Everything you need to know about AI (for PMs and builders)

* Carl Vellotti on Claude Code

* Marily Nika on Google AI PM Tool Stack

* Pawel Huryn on n8n for PMs

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

Today's Episode

Carl Vellotti is back.

His first episode hit 30k views. My most popular episode in months.

Today, he’s giving you the advanced playbook for Claude Code.

Claude Code hit $1 billion ARR in December. Fastest product ever to reach that milestone. Under 6 months.

Carl has been building with Claude Code every single day. He’s made all the mistakes. He’s learned all the expert workflows and tips.

In today’s episode, he shares everything.

----

If you want access to my AI tool stack - Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin - for free, grab Aakash’s bundle.

Are you searching for a PM job? Join me + 29 others for an intensive 12-week experience to master getting a PM job.

----

Brought to you by:

Amplitude: The market-leader in product analytics

Pendo: the #1 Software Experience Management Platform

Jira Product Discovery: Move discovery and roadmapping out of spreadsheets

Miro: The innovation workspace

----

Key Takeaways

1. Claude Code hit $1B ARR in 6 months - fastest product ever. Anthropic optimized for power users doing deep work. Fewer users, insanely deep usage. Carl: "I spend all day in Claude Code. It gets to the point where you never really have to leave."

2. MCPs connect Claude to everything in 2 minutes - Model Context Protocol invented by Anthropic. One command to install Linear, Google Workspace, Slack. Essential stack: documents first, task management second, communication third, data sources fourth.

3. End-to-end PM workflow takes one morning instead of one week - Survey creation → analysis → PRD → presentation → 19 tickets. All through MCPs. Carl didn't use templates. Opus 4.5 handled everything.

4. Skills are pre-built templates but don't always auto-trigger yet - Solution: specify explicitly. "Create presentation using your presentation skill." Build Skills for anything you do more than twice. That's the compound effect.

5. Real Google Slides, not images - Carl's presentation skill created 19 fully editable slides from PRD. Takes a few minutes because Claude creates slides in parallel. "This is saving you hours. We used to spend entire afternoons turning PRDs into decks."

6. GitHub turns Claude into remote worker - Create issues from your phone. @mention Claude. It works while you're away. Carl's journal is a private GitHub repo. Voice transcriptions → issues → Claude updates markdown.

7. Workflows beat agents for production - Level 1 workflow: 5,000 tokens, 40 seconds. Level 3 agent: 90,000 tokens, 90 seconds. Carl's rule: "If something can be expressed as code, code it. Leave agents for cognitive work."

8. Production best practices separate hobbyists from pros - Set error workflows. Add retry logic (3 attempts, 1-second delays). Pin data during development. Use reasoning effort appropriately. Write detailed tool descriptions.

9. Real challenge is building reusable workflows - Learning basics takes a day. First workflow takes hours. But having 50 workflows built over months means everything is automated. Carl: "Because I have workflows for all these things already built, I can do it all in one shot."

10. Tools improve faster than you can learn them - Carl: "By the time you see this, there might be Claude 5." Your workflows from today will 10x when better models drop. PMs who started 6 months ago aren't just 6 months ahead, they're 100x ahead because systems compound with every improvement.

----

Where to Find Carl Vellotti

* Linkedin

* X (Twitter)

* Instagram

* Newsletter

Related Content

Newsletters:

* How to Use Claude Code Like a Pro

* Steal 6 of My Claude Skills

* Claude Skills Tutorial

* AI Agents: The Ultimate Guide for PMs

* Practical AI Agents

Podcasts:

* The Ultimate Guide to n8n for PMs, with Pawel Huryn

* This is what a Google AI PM’s Tool Stack Looks Like, with Marily Nika

* How to Use Claude Code Like a Pro (Carl’s First Episode)

----

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

If you want to advertise, email productgrowthppp at gmail.

Marily Nika and I filmed my 7th most popular episode a year ago.

She broke down everything on how to become an AIPM because she’s done it. She is an AIPM at Google with 11 years of experience.

Today, she’s back with a completely new episode. She’s dropping every AI tool she uses daily as an AIPM.

Not theory. Not hype. The actual 6 tools she uses multiple times every single day.

Today she’s dropping all the expert tips and tricks. She’s made the mistakes so you don’t have to.----Today’s guide covers:

* The AI PM Tool Stack

* Google AI Studio for Prototyping

* Opal for Mini Apps

* Notebook LM for Domain Expertise

* Perplexity for Reddit Research

* ChatGPT for Your Voice

* Fireflies and Getting Tools at Work

* The 18-Month AIPM Roadmap

* The AIPM Interview Red Flags

----Check out the conversation on Apple, Spotify and YouTube.

Brought to you by:

* Jira Product Discovery: Move discovery and roadmapping out of spreadsheets

* Vanta: Automate compliance across 35+ frameworks

* Maven: Improve your PM skills with awesome courses. Discount with my link

* Mobbin: Discover real-world design inspiration

* Product Faculty: Get $500 off the AI PM Certification with code AAKASH25----

Key Takeaways

1. Marily's stack is just 6 tools - Google AI Studio for prototyping, Opal for mini apps, Notebook LM for learning, Perplexity for user research, ChatGPT for PRDs, and Fireflies for meetings. Each maps to one PM workflow. No tool hopping.

2. Prototype before writing PRDs. Marily builds a working app in AI Studio first, then brings engineers in to debate the actual thing. Saves weeks of back-and-forth on documents. PRDs are now for complex cross-functional work only.

3. Notebook LM learned a 4-hour video in 15 minutes. Marily had an interview the next day. She uploaded the job description and a 4-hour investor relations video. It gave her 15 key points. She memorized them and crushed the interview.

4. Perplexity's Reddit filter is your secret weapon. Turn off web search, turn on discussions and opinions. Ask "would young professionals be interested in a fitness ring?" Get 20+ Reddit sources instantly. Know what users actually want.

5. Marily's PRD generator has 10,000+ users. She trained a ChatGPT custom GPT with her voice and old PRDs. It asks probing questions before generating. Gets PRDs done in days instead of weeks. She's not embarrassed to use AI at work.

6. Tool selection has 4 rules. Does it save 10x time (not 2x)? Does it work across contexts? Does it work with your company's limitations? Does it compound over time? If not all 4, delete it.

7. "Be like a crab" for AIPM roles. Move adjacent to your experience. Hearing aids → AirPods PM. ESPN journalism → Meta sports AI PM. Don't ignore your past experience. It's your competitive advantage.

8. Red flag in AIPM interviews - missing PM craft. Ex-ML scientists dive into algorithms without asking why, who, and how to measure success. AI changes HOW you fulfill use cases, not the use cases themselves.

9. Who replaces PMs? Not AI. PMs who use AI when you're not. Don't be embarrassed about using AI at work. It's not cheating. The gap is widening between PMs who use these tools and those who don't.

10. Every PM will become an AIPM by 2026. Show Marily any product and she'll find the AI use case. Even retail stores use cameras and AI to monitor which areas get traffic. The AIPM title will just become PM.

----

Where to Find Marily

* LinkedIn

* Twitter

* Website

* YouTube

* Newsletter

* Maven - Her course is part of my special curation with Maven

----

Related Content

Newsletters

* AI Foundations for PMs

* How to Become an AIPM

* AI Prototyping for PMs

* AI PRDs

* The Complete AI PM Course

* AI Product Strategy

* ChatGPT for PMs

Podcasts

* Pawel Huryn on n8n

* Alex Danilowicz on AI Prototyping

* Hamza Farooq on Landing AIPM Jobs

* Todd Olson on Upskilling to AIPM

* Rachel Wolan on AI Product Leadership----

If you want access to AI tools for free, get Aakash’s bundle and access to 9 AI products for an entire year.

Are you searching for a PM job? Join me + 29 others for an intensive 12-week experience to master getting a PM job.

----

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!If you want to advertise, email productgrowthppp at gmail.

Pawel Huryn is the guest behind my most popular episode ever (52K+ views).

Today, he’s back to give you a masterclass in one of the most exciting AI tools out there: n8n.

n8n is the most powerful workflow automation tool that combines two things: traditional workflow automation and building AI agents.

And Pawel has been knee-deep in n8n more than almost anyone else in the world.

He’s tried everything. He’s made all the mistakes. He’s learned all the expert workflows and tips and tricks.

In today’s episode, he walks through building real n8n workflows from scratch.

-----

Check out the conversation on Apple, Spotify and YouTube.

Brought to you by:

* Amplitude: The market-leader in product analytics

* Vanta: Automate compliance across 35+ frameworks

* Testkube: Leading test orchestration platform

* Kameleoon: AI experimentation platform

* Pendo: The #1 software experience management platform

-----

Key Takeaways

1. N8N combines traditional workflow automation AND AI agent building in one platform - making it more powerful than Zapier or Make for complex automation needs.

2. Real use cases span from simple business workflows to chatbots, automatic competitor monitoring, multi-agent research systems, and inbox workers that take actions based on emails. Sky is the limit.

3. Pavel's competitor monitoring workflow costs $1-2/week using the FREE version of N8N. Just needs Perplexity API ($1-2 for hundreds of calls) and OpenAI credits. Enterprise tools charge $500+/month.

4. Pin your data during development. N8N caches API responses so you don't burn credits while testing workflows. Click the pin icon and N8N uses cached data instead of making new API calls.

5. N8N automatically loops through items - no need to write for-loops or while-loops. When you connect a node with 6 items, N8N repeats the action 6 times automatically.

6. Compress context before sending to LLMs. Pavel cuts 70% of tokens by extracting only summary content and citation URLs from Perplexity results, ignoring irrelevant snippets and metadata.

7. Use ChatGPT to write N8N code snippets. Pavel never writes code blocks himself - just takes a screenshot of the data and asks GPT "how do I compress this information?"

8. Traditional workflows are more efficient (saves tokens, very reliable) for predictable tasks. AI agents are more flexible but use more tokens and can make mistakes. Use workflows when you know the steps.

9. Set GPT reasoning effort to "low" for simple tasks. When you just need formatting or summarization (not complex thinking), low reasoning effort saves tokens significantly.

10. Best practices: Set dedicated error probes to catch errors before they break workflows. Use max iterations to prevent infinite loops. Set retry on fail to 3x attempts. Pin data during development.

-----

Where to Find Pawel

* LinkedIn

* Newsletter

* YouTube

-----

Related Content

Newsletters:

* Ultimate Guide to AI Prototyping Tools

* AI Agents: The Ultimate Guide for PMs

* Practical AI Agents

Podcasts:

* AI Agents Demo, with CEO of Relay.app

* How to Build AI Agents (and Get Paid $750K+)

* 5 AI Agents Every PM Should Build, with CEO of Lindy

If you want access to a n8n competitor for free, get Aakash’s bundle and access to Relay.app.

-----

If you want to advertise, email productgrowthppp at gmail.

Today’s Episode

Laura Burkhauser started as an IC PM at Descript, the $550M AI video editing platform.

Three years later, she’s CEO.

She shipped AI features that worked.

In today’s episode, you’ll hear the exact features Laura built, the eval framework she used, and her complete IC → CEO path.

This might be the most actionable career episode you listen to all year.

----

Check out the conversation on Apple, Spotify and YouTube.

Brought to you by:

* Maven: Improve your PM skills with awesome courses

* Pendo: #1 Software Experience Management Platform

* Vanta: Automate compliance across 35+ frameworks like SOC 2 and ISO 27001

* NayaOne: Airgapped cloud-agnostic sandbox

* Kameleoon: Leading AI experimentation platform

----

Key Takeaways1. Map your user journey BEFORE picking AI features. DeScript identified pain points (retakes, eye contact, rambling), then asked "what just became possible with LLMs?" Build that intersection.2. Build prepackaged buttons, not blank chat boxes. Each DeScript AI tool is a carefully crafted prompt behind a single button that delivers reliable results every time.3. Use human evals on production data before shipping. Test on real customer data, ask "would I use this as a customer?" If yes, ship. If no, don't.4. The ultimate metric is export rate. If users apply your AI feature then remove it before exporting, it didn't meet their quality bar.5. Switch from buttons to chat when you hit 30+ parameters. When users wanted topic selection, speaker choice, and platform optimization, chat became better than buttons.6. Match your eval data to actual use case. DeScript failed with Studio Sound because they tested on terrible audio (vacuuming, jackhammers) when real users had laptop microphones. Different models handle different quality levels.7. Test agents with real customer language early. Don't use toy data or employee terminology. Mix sophistication levels—some advanced at video and AI, some complete beginners—to understand how real people prompt.8. Launch AI agents to new users first. Video editing is hard and many people quit. DeScript tested Underlord on activation and it won, so new users got it first before existing users.9. Choose breadth over depth for product-wide agents. DeScript chose breadth—Underlord works across all features because "we're not a point solution." Requires more context, tool coverage, and evals but serves the product vision.10. Earn founder trust by getting command, not by being strategic. Use the product extensively. Talk to customers constantly. When you speak, people think "Smart" and invite you to more rooms. Ship features before focusing on strategy.

----

Where to Find Laura Burkhauser

* LinkedIn

* Company

----

Related Content

Podcasts:

* AI Product Leadership Masterclass

* Conversation with the CEO and Founder of Bolt

* This $20M AI Founder Is Challenging Elon and Sam Altman | Roy Lee, Cluely

Newsletters:

* Ultimate Guide to AI Prototyping Tools

* AI Agents: The Ultimate Guide for PMs

* How to Land a $300K+ AI Product Manager Job

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

Check out the conversation on Apple, Spotify and YouTube.

Brought to you by:

* Jira Product Discovery: Move discovery and roadmapping out of spreadsheets

* Vanta: Automate compliance across 35+ frameworks like SOC 2 and ISO 27001

* Kameleoon: Leading AI experimentation platform

* NayaOne: Airgapped cloud-agnostic sandbox

* Product Faculty: Get $500 off the AI PM Certification with code AAKASH25

Today’s Episode

Claude Code hit $1B ARR in 6 months. But OpenAI is not just giving up.

ChatGPT’s new Codex is the most powerful way for product managers to build prototypes.

And it’s a far better way to use ChatGPT than the browser.

So every PM should know how to use it.

Today, I brought back the man behind my Claude Code tutorial, Carl Vellotti, for a full guide on how to use ChatGPT Codex for PMs:

This might be the most important podcast you watch all year.

(And not a single other PM podcast has even talked about this tool)

Your Newsletter Subscriber Bonus

For subscribers, each episode I also write up a newsletter version of the podcast. Thank you for having me in your inbox.

Today’s guide covers:

* Getting Started: The Non-Technical PM’s Guide to Codex CLI

* Classic PM Tasks: Documents, Meetings & PRDs with Codex

* Advanced Prototyping: From Vibe Coder to Vibe Engineer

This is your complete Codex roadmap.

1. How to Get Started - The Non-Technical PM’s Guide to Codex CLI

If you’ve never touched a terminal, this is for you.

1a. Open Codex in an IDE

Carl opens Cursor and shows files Codex creates in real-time.

Workflow: Open project folder in Cursor → Terminal (Control + backtick) → Type codex.

Now you see files created, preview documents, navigate visually.

As Carl Says:

“Open it in an IDE. Easiest way to see what it’s actually doing.”

1b. The YOLO Mode Hack

Codex asks for permission constantly—every website, every command.

Solution:

codex --yolo

Full access mode. No prompts. Just execution.

“Haven’t broken my computer yet. We’re in this directory so it won’t leave.”

1c. Codex vs Claude Code

Same task on both: Search web, summarize differences.

Claude Code: 3 searches simultaneously, ~2 minCodex: One site at a time, asks permissions, ~4 min

Claude Code is faster, hides details. Codex shows every command—more verbose.

“It’s Apple versus Microsoft. Claude does things in a nicer interface. Codex shows you everything.”

2. How to Handle Classic PM Tasks with Codex

Codex shines at daily PM work.

2a. File Analysis Without Uploading

Run Codex from a folder, it accesses everything. No uploading.

Carl has demo folder (Taskflow) with interviews, notes, PRDs.

“What user interviews completed?” → Codex lists them“Top 3 pain points?” → Returns: voice input reliability, integration gaps, template workflow

Creates document with direct quotes—no manual file providing.

2b. Template-Driven Workflows

Solution: Create /templates folder with markdown files.

Example: Discussion Points, Action Items, Risks & Blockers, Next Steps

Use: Summarize @meeting-notes using @template

Same format every time. Works for PRDs, one-pagers, research summaries.

2c. Socratic Questioning for Better PRDs

Problem: Asking Codex to write a PRD produces garbage without proper thinking.

Solution: Socratic template makes Codex ask YOU questions first.

Questions: “Why is this helpful? Data, feedback, or strategic?” “What must work for V1?” “Edge cases?”

You answer. AI embeds your thinking.

Then Codex reviews context, templates, example PRDs, and writes.

Carl: “Goes from kind of there to a really good PRD almost right out of the box.”

3. Advanced Codex Techniques for Future Vibe Engineers

This is where Codex separates from the browser.

3a. Design Systems with Storybook

Vibe coder problem: “Make title pink” → wrong shade → 20 iterations → still janky

Vibe engineer solution: Storybook (npm run storybook)

See components visually, change colors instantly, see changes live without redeploying.

Carl changes recipe title to pink: read-only mode to see plan, then YOLO mode to execute. Updates in Storybook immediately.

Pre-built components: Use Shadcn UI for production-grade React components. Calendar, date picker, dropdown—all done.

“All the logic and difficult things, you get for free.”

3b. Test-Driven Development is The Unlock

Problem: AI says “Done!” but it’s broken. AI will lie and set variables to true to pass tests.

Solution: Write tests BEFORE building.

Red-Green-Refactor: Tests fail → Build until pass → Improve code

Carl’s macro calculator:

Tests: API returns calories, empty recipe returns null, missing data shows “N/A”, API fails retries 3x, divide by servings correctly

Told Codex: “Build to make tests pass.”

35 minutes. Zero touch. All tests green. Feature worked.

Carl: “If you write tests FIRST, they can’t cheat.”

3c. An Example - The TikTok Recipe Bot

Carl built this for his girlfriend: TikTok recipe extractor.

Problem: Recipe videos on TikTok. No written recipe. Just “comment and I’ll DM.”

Solution: Paste URL → Codex downloads video → Sends to Gemini (only model that processes video) → Gemini extracts recipe → Codex formats to PDF

How: Detailed implementation spec: architecture, data flow, APIs, error handling, retry logic, tests.

Gave to Codex in GPT-5 Codex mode. Left 35 minutes.

Built entire feature. Tests passed. Worked first try.

Carl:

“First time I saw the dream—give it a medium-sized feature and it just builds it.”

Where to Find Carl Vellotti

* Linkedin

* X (Twitter)

* Instagram

* Newsletter

Related Content

Podcasts:

* Claude Code Tutorial

* Windsurf Tutorial

* AI Prototyping Tutorial

Newsletters:

* Ultimate Guide to AI Prototyping Tools

* AI Agents: The Ultimate Guide for PMs

* How to Land a $300K+ AI Product Manager Job

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

Today’s Episode

AI PM jobs pay 30-40% more than regular PM jobs.

But here’s the problem: You can’t just slap “AI PM” on your resume.

Todd Olson has spent 28 years in product management, VP of Product at a public company, then founder of Pendo, now a $2.5B product management platform working with everyone from American Cancer Society to Zendesk.

----

Check out the conversation on Apple, Spotify and YouTube.

Brought to you by - Reforge:

Get 1 month free of Reforge Build (the AI prototyping tool built for PMs) with code BUILD

----

Key Takeaways

1. AI PM market exploded - Last year 10% of PM jobs were AI PM jobs. This year it's 20%. They pay 30-40% more because of scarcity and skill level. But Todd warns: "You better damn well be good and know what you're talking about if you're gonna call yourself an AI PM because we are going to interrogate the hell out of it."2. Real requirement is production at scale - Not "I built prototype at 1-person startup." Hiring managers want 20,000 paying B2B customers experiencing your AI feature successfully. To get there: upskill internally at current company by shipping AI features on your roadmap.3. The 5-layer technical pyramid - Foundation: AI/ML fundamentals, data pipelines, prompt engineering. Middle: Observability (trace analysis), cost optimization, evals. Top: Product strategy, stakeholder management, leadership. You need to climb all 5 layers. Most PMs stop at layer 1.4. RAG is table stakes - "RAG is the de facto way to build." You ingest data, create embeddings, feed into vector database, look up relevant context, pass to LLM. Todd: "If you put too much in context window, just like a human, you get confused. You want to give the right context."5. PM-engineering tension is real - At startups, PMs do trace analysis. At large companies, engineering managers push back: "This is my world. I don't want some PM shadowing me." Similar to Data Dog—most PMs don't have login. Know the line. Be fluent but respect boundaries.6. But evals are YOUR domain - Unlike trace analysis, evals are where PMs are the expert. "The PM is probably the best-suited human being to author and manage eval sets." You understand user and business needs. Engineers don't have that context. This is must-have competency now.7. Cost optimization will matter - Some AI companies have sub-15% gross margins. Traditional software is 70-80%. Todd: "It's not a business at sub-15%." Eventually you'll rearchitect systems because infrastructure is too costly. Rule: when something's faster, it's cheaper (both buying compute).8. Solve hard problems, not shiny objects - Todd's test: "Are we gonna do much better job than ChatGPT out of box? Why would we just wrap that and slap Pendo logo on it?" His discovery agent example: hard part isn't interviewing customers—it's finding which to interview, prioritizing, scheduling. Automate that workflow.9. Kill bad features ruthlessly - Todd shipped features couple years ago that weren't great and turned them off. "Too often we hold on to something. Turn them off. Be unafraid. The more stuff in your product, the worse the experience is by default."10. Control the narrative with boards - Don't show up with no story and get crushed with random requests. Todd: "Show them how you actually run your business. I want to see what you're looking at, not something just made for me." Think deeply about how each bet drives shareholder value.

----

Where to Find Todd Olson

* LinkedIn

* Company

* X

----

Related Content

Podcasts:

* How to Become, and Succeed as, an AI PM | The Marily Nika Episode

* If you only have 2 hrs, this is how to become an AI PM

* Complete Course: AI Product Management

Newsletters:

* How to Become an AI Product Manager with No Experience

* How to Write a Killer AI Product Manager Resume

* How to become an AI Product Manager

----

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

----

If you want to advertise, email productgrowthppp at gmail.

Today’s Episode

There are tons of tutorials about Claude Code and Cursor for IC PMs.

But what about leaders?

Today’s episode is a masterclass on both sides of AI product leadership. How to be a productive AI leader and how to ship AI-native features at scale.

Rachel Wolan is the Chief Product Officer at Webflow, the $4 billion company powering TED Talks, SoundCloud, and Reddit.

Rachel walks through building her agentic Chief of Staff live, sets up a LinkedIn post generator from scratch, and shares the brutal lessons from launching Webflow’s AI app generator.

----

Brought to you by:

Linear: The task management platform dethroning Jira

-----

Key Takeaways:

1. IC CPO means self-serving answers - "As a leader, you are able to get your own answers to practically any question." No waiting on data scientists. No back-and-forth with analytics. You have tools to self-serve insights, make analysis, automate workflows. Model behavior for your team to inspire them.

2. Calendar agent analyzes time - Runs weekly with prompt: "Analyze my calendar for last two weeks. Where could I delegate?" Returns delegation opportunities, red flags (double bookings, context switching), what to cut next week. Rachel gives output to EA. Spot on when shown live.

3. Email agent watches behavior - Complete inbox access. Runs triage, archives junk (calendar notifications, marketing), pins important messages, creates draft replies. Twist: watches behavior. If email sits too long, it notices. Caught meeting missing link. Rachel's rule: agent recommends, she approves. No autonomous sending.

4. Analytics agent via MCP - Connected Claude Code to Snowflake via MCP servers (not officially supported repos, just fed them to Claude Code). Ask natural language questions, get SQL executed real-time. "How many sites does Shirts.com have?" Claude writes query, authenticates via SSO, returns answer. Data scientist in pocket.

5. Accept the adoption curve - Your org follows standard curve: early adopters, early majority, late adopters, laggards. Create pathways for everyone to ascend ladder at their pace. Don't force everyone to be you. Rachel to team: "I only want to see prototypes when you have meetings with me." Creates culture investing in prototype quality.

6. Builder Days strategy - Give everyone access: Claude Code licenses, MCP to Snowflake/Tableau, Figma Make, Cursor with design system. Run Builder Days where champions help others through technical hurdles. Everyone demos something outside comfort zone. Results: 0% to 30% of designers using Cursor weekly after first Design Builder Day.

7. Rewrite career ladder - Webflow rewriting career ladder to make AI-native work an expectation, not nice-to-have. Create right incentives. Make sure people supported. Avoid AI for AI's sake. Example: Two designers built similar prototypes. Director caught early: "Go harmonize your prototypes now." Easier now than late in product cycle.

8. MVO before MVP framework - Most teams: Feature → PRD → Design → Ship. Rachel flips it. MVO (Minimal Viable Output) before MVP. Get model's output right FIRST using RAG, prompt engineering, context engineering. Only then build feature. "If you don't have desired outputs, don't spend time productizing the AI feature."

9. Evals are now your job - Brutal story: Webflow's AI app generator 2 weeks from launch. Rachel tested it. Agent kept dying. Realized: changed underlying model, evals didn't have coverage. Evals = test cases for models. Want dream evals (should pass) and edge cases (should fail). Use BrainTrust. Teaching PMs to write evals is part of AI PM toolkit now.

10. Build on your strengths - Framework: See trend → Is it applicable to customers? → What's YOUR core competency? Webflow's strength: bringing visitors to front door via CMS. Built production-grade app generator (not prototype like Lovable). Uses your brand, CMS, hosting, security. "We're bringing a way to prompt an app to production." Don't copy trends, leverage unique strengths.

-----

Where to Find Rachel Wolan

* LinkedIn

* Website

* X

----

Related Content

* Claude Code Tutorial for AI PMs

* AI Agents for PMs in 69 Minutes, with IBM VP

* 5 AI Agents Every PM Should Build, with CEO of Lindy

Newsletters:

* AI Evals Guide for PMs

* Prompt Engineering for AI Agents

* AI Agents: The Ultimate Guide for PMs

----

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps

----

If you want to advertise, email productgrowthppp at gmail.

Today’s Episode

Why do your prompts keep failing?

You write the perfect prompt. The AI spits out garbage. You tweak. You iterate. You spend hours getting mediocre results.

XK built Cues to $10M ARR in 60 days with zero VC funding and zero advertising. Today, he’s dropping the complete playbook:

-----

Check out the conversation on Apple, Spotify and YouTube.

Brought to you by:

Reforge http://reforge.com/aakash

-----

Key takeaways:

1. Context engineering beats prompting - One prompt won't work. Like hiring someone who knows nothing about your company—impossible to get results in 5 seconds. Accumulate context, build knowledge base, let AI know you over time. Combines system prompts, user prompts, memory, and RAG.2. The Mom analogy - Your mom knows your preferences, goals (grow taller for basketball), what makes you happy. She doesn't need detailed instructions. That's context engineering. AI that knows you creates better results and positive loops.3. Threads growth hack - Created hundreds of accounts posting use cases daily. Zero ad spend. Why it works: Threads gives traffic generously, less crowded than X, no creator hierarchy. Result: 3M impressions/month, hundreds of daily visits. Targeted Taiwan/Hong Kong markets.4. MVO before MVP - Traditional: Feature → PRD → Design → Ship. Xiankun's way: Get model output right FIRST. Use RAG, prompting, fine-tuning for Minimal Viable Output. Then productize. "If no desired outputs, don't spend time productizing."5. Visual context engineering - Use spatial tools: draw squares, graphs, sketches. AI understands spatial relationships. Unlike ChatGPT where files disappear, Kuse gives 2D space to store/reuse. Graphic operating system for AI that compounds.6. The pivot story - Started as design agent. Users uploaded documents instead. Knowledge base usage far exceeded design. Pivoted to horizontal knowledge-based AI. Listen to your users.7. Why X sucks for growth - Structured creator hierarchy. Can't farm traffic without famous connections. Good for VC fundraising, terrible for user acquisition. Threads and Instagram are underserved with real users.8. Compounding context power - Regular chatbots: one-off, context disappears. Kuse: processes files when you're away, pre-prepares everything. Like having ingredients ready vs ordering each time. Each interaction improves.9. Trading company origin - Co-founded YC company, created trading company, made money, funded Kuse with profits. Built without VC pressure. "Entrepreneurship is a game of focus." Building without chasing VC gives fresh perspective.10. Future vision: productivity playground - "Not building productivity tool, building playground." When AI takes jobs (2030-2040), people need fulfillment. Kuse is amusement park where people pretend to work, feel satisfaction. Going to pure pleasure, not efficiency.

----

Where to Find Xiankun Wu

* LinkedIn

* Threads

* Company

----

Related Content

Podcasts:

* We Built an AI Employee in 62 mins

* Conversation with the CEO and Founder of Bolt

* This $20M AI Founder Is Challenging Elon and Sam Altman | Roy Lee, Cluely

Newsletters:

* Context Engineering Guide

* Prompt Engineering in 2025

* How to become an AI Product Manager

----

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

----

If you want to advertise, email productgrowthppp at gmail.