Paweł Huryn

Just shared PM Skills v2.0 with 9 plugins, 68 skills, 42 commands. Free, MIT.

You get the rigor of Teresa Torres, Marty Cagan, and Alberto Savoia built into your daily workflows.

The open-source repo (12K GitHub stars) is now bigger, and it reaches further:

→ /red-team-prd: attacks the plan before the review does

→ /ship-check: documents the system, audits the code against the intent, maps test coverage, and audits performance and security

→ The 8 previous plugins, such as pm-product-discovery, pm-execution, pm-data-analytics, and pm-strategy are still supported

→ Cowork and Claude Code, MIT-licensed


Install it and run /red-team-prd on the plan you're least sure about. That's the fastest way to see the difference.

I used /ship-check to responsibly disclose bugs in public repos, such as Langfuse (already fixed).


GitHub: https://lnkd.in/dxdTj24P

Full instructions (+subscribe free for updates): https://lnkd.in/dhsAM3yk

Today, my vibe product hit 100K installs. 37K MAU.
~177K lines of code I haven't even looked at.

Anyone can do the same. And stay safe.

As a PM, you don't have to code. Just understand engineering enough to know what questions to ask AI.

Free resources to get you started below. I wish I had them before.

[Day 1] Review these helicopter view resources:

- System Design 101 by ByteByteGo: https://lnkd.in/dcaXetER
- Engineering Visual Guides: https://lnkd.in/dgvvU_EZ
- Become a Supabase Pro in 1.5 Hours (1:26:39): https://lnkd.in/dGJqThnT
- Overview of HTTP by Mozilla: https://lnkd.in/dzSGWpyy
- HTTP Request Methods: https://lnkd.in/d_sVPjDG
- The most important concepts of Node.js and Express.js (17:09): https://lnkd.in/dFPvPybP
- Every React Concept Explained in 12 Minutes (11:52): https://lnkd.in/dhXMPdQ3
- TypeScript - The Basics (12:00): https://lnkd.in/d26iPzem
- CSP Guide by Mozilla: https://lnkd.in/dNMhZD9v
- OWASP Top 10: https://lnkd.in/dRPY7dGu
- CI/CD by Netlify: https://lnkd.in/dGvrCmiX
- Overview of Test Automation with Selenium: https://www.selenium.dev/

(you will get back to them later)

[Days 2-3] Start working with Claude Code and Codex (free 1h+ recordings with templates):

- Codex for PMs: https://lnkd.in/dqaRBzdu
- Claude Code for PMs: https://lnkd.in/dqaRBzdu

[Days 4] Proven ways to manage AI agents (not what people say):

- Intent Engineering: https://lnkd.in/dhBzF-zB
- Don't review code, review artifacts: https://lnkd.in/daw-Nt-m

[Day 5] Continue building

Install 5 skills from https://lnkd.in/dxdTj24P (25K GitHub stars):
- /pm-ai-shipping:ship-check
- /pm-ai-shipping:document-app
- /pm-ai-shipping:derive-tests
- /pm-ai-shipping:security-audit-static
- /pm-ai-shipping:performance-audit-static

See what's inside. Together, they greenlight the release.

-----

P.S. Looking for a more structured program?

8 AI PM cohorts (Anthropic, Google, OpenAI) in the price of one + $1,500 off only for my community: 👉 https://lnkd.in/dS6sgqnZ

Plus, everyone who enrolls by Sep 5 can attend my Claudathon for PMs for free: https://lnkd.in/dWNfApWB (send me an email with a purchase confirmation: pawel@productcompass.pm)

-----

*My product: "Grok Build for VS Code." 90% of users came from the word-of-mouth, not social media.

- Open VSX (77K installs): https://lnkd.in/dwt7aUgK
- VS Code marketplace (23K installs): https://lnkd.in/dc6emNZm
- Desktop app (10K installs): https://lnkd.in/dcKqbSeB

37K MAU. 5,200+ unit tests planned and written by the agents.

I open-sourced the second brain for product managers.

Sibling repo to pm-skills (11K stars on GitHub). Memory layer to its workflow layer.

Folder of markdown files in a git repo. The agent reads them before answering, writes after, sweeps weekly.

No vector DB. No embeddings. Grep-able.

What's inside:

→ knowledge/: your durable picture of strategy, product, users, market
→ hypotheses/: what you're testing, with evidence trails
→ decisions/: every shipped + killed option, dated, with rationale
→ interviews/: raw transcripts tagged by persona, stage, risk
→ sweeps/: weekly maintenance that promotes durable patterns up

Every claim is tagged. Observation, interpretation, hypothesis, assumption, decision. A Slack comment is not automatically truth.

I designed 17 PM scenarios from what I've shipped through. Ran them through an eval harness. 404 of 406 individual checks pass on Claude Sonnet 4.6.

Star it, fork it, fill it with your own context.

Free repo: https://lnkd.in/dBN-vWhN

Full article & new PM story: https://lnkd.in/dbyj5pbz

Steal my 100+ agentic skills, commands, and plugins for PMs. Designed for Claude Code & Cowork. It's the AI OS for every PM 👇

Each skill encodes a proven PM framework and walks you through it step by step. You get the rigor of Teresa Torres, Marty Cagan, and Alberto Savoia built into your daily workflow.

Inside 8 open source plugins:


1. Product Discovery

Ideation, experiments, assumption testing, Opportunity Solution Trees, customer interviews, and feature prioritization

2. Product Strategy

Vision, strategy, business models, pricing, SWOT, PESTLE, Porter's Five Forces, Ansoff Matrix, and competitive landscape

3. Execution

PRDs, OKRs, roadmaps, sprints, retros, release notes, pre-mortems, stakeholder maps, and user stories

4. Market Research

Personas, segmentation, journey maps, market sizing (TAM/SAM/SOM), and competitor analysis

5. Data Analytics

SQL query generation, cohort analysis, retention curves, and A/B test statistical analysis

6. Go-to-Market

GTM strategy, beachhead segments, ICP, growth loops, GTM motions, and competitive battlecards

7. Marketing & Growth

Marketing ideas, positioning, value prop statements, product naming, and North Star metrics

8. PM Toolkit

Resume review & tailoring, NDA drafting, privacy policies, and grammar/flow checking


Full catalog + how to install in a few clicks on GitHub: https://lnkd.in/dxdTj24P

If this helps you, ⭐ the repo!


By the way, if you want to break into AI PM, I recommend the AI PM Certification. A special $500 discount + 1-year subscription to my newsletter only with this link: https://bit.ly/aipmcohort

Claude Cowork now runs any LLM. You can test it free.

OpenAI, Gemma, Kimi K2, local models, or via enterprise gateway (Bedrock, Vertex, Foundry).

Anthropic shipped it quietly. I discovered it by accident.

OpenRouter CEO confirmed on X they support it, too.

---

Setup in 5 minutes:

1. Menu → Developer → Configure Third-Party Inference
2. Connection: Gateway
3. Gateway base URL: https://openrouter.ai/api
4. Gateway API key: your OpenRouter key
5. Set "Allowed egress hosts" under Sandbox & workspace. My config: "*"
6. [Apply locally] → [Relaunch]

You don't need a Claude subscription.

---

What's missing out of the box:

→ Skills: upload works, list is empty. Install the official ones (docx, pdf, pptx, xlsx, skill-creator): https://lnkd.in/dquQZjra

→ Plugins: upload works, empty. Pull PM, legal, finance, or marketing from: https://lnkd.in/d6ZJbET3

→ Connectors: no built-in MCPs. Add them manually. A great collection: https://lnkd.in/dnaUyRg6

---

The real story is the enterprise config:

→ Max tokens per window (per-user soft cap)
→ Allow user-added MCP servers
→ OpenTelemetry collector endpoint
→ Block auto-updates
→ Remove built-in tools from Cowork

This isn't a consumer feature. It's Claude Desktop becoming a managed agent platform.

The direction is clear: Anthropic wants the harness to be the default, regardless of which model sits behind it.

---

A full free breakdown with step-by-step setup, every gotcha, and troubleshooting: https://lnkd.in/d8YTXJNK

12-Month MBA for PMs
(better than any $135K program)

👨‍🎓 Month 1:

- Inspired by Marty Cagan
- Empowered by Cagan and Chris Jones
- The Right It by Alberto Savoia

👨‍🎓 Month 2:

- Continuous Discovery Habits by Teresa Torres
- The Mom Test by Rob Fitzpatrick
- The Lean Product Playbook by Dan Olsen

👨‍🎓 Month 3:

- Business Model Generation by Alexander Osterwalder
- Running Lean by Ash Maurya
- Testing Business Ideas by Strategyzer

👨‍🎓 Month 4:

- Radical Focus by Christina Wodtke
- No Rules Rules by Reed Hastings
- Team Topologies by Manuel Pais and Matthew Skelton

👨‍🎓 Month 5:

- Product-Led Growth by Wes Bush
- The Product-Led Playbook by Wes Bush
- Hooked by Nir Eyal

👨‍🎓 Month 6:

- This is Marketing by Seth Godin
- GTM Strategist by Maja Voje
- Hacking Growth by Sean Ellis

👨‍🎓 Month 7:

- Talk Like TED by Carmine Gallo
- The Copywriter's Handbook by Robert Bly
- Never Split the Difference by Chris Voss

👨‍🎓 Month 8:

- Thinking, Fast and Slow by Daniel Kahneman
- The 7 Habits of Highly Effective People
- Superconnector by Paugh and Gerber

👨‍🎓 Month 9:

- The Making of a Manager by Julie Zhuo
- The Psychology of Money by Morgan Housel
- Accounting for Non-Accountants by Wayne Label

👨‍🎓 Month 10:

- Lean Analytics by Alistair Croll and Ben Yoskovitz
- North Star Framework Field Guide (free PDF): https://lnkd.in/djvDpyVW
- Are You Tracking The Right Metrics (free): https://lnkd.in/d9Mw3Bp8

👨‍🎓 Month 11:

- Jobs-to-be-Done by Tony Ulwick
- Escaping the Build Trap by Melissa Perri
- Build by Tony Fadell

👨‍🎓 Month 12:

- Playing to Win by Roger Martin
- The Invincible Company by Strategyzer

-

No AI books on the list. On purpose.

You learn AI by doing, not by reading. 7 steps, 10 minutes, no coding: https://lnkd.in/dE6tuFXa

-

Bonus (recommended):

- Getting a PM Job by Aakash Gupta
- Working Backwards by Bryar and Carr
- The MBA Pathology by Marty Cagan: https://lnkd.in/dw6kpwTv
- 650+ free PM learning resources in the welcome email: https://lnkd.in/d7g76cqZ

-

All it takes is 45-60 min/day of reading or audiobooks.

P.S. You can network for free on LinkedIn.

Most PMs run Claude the most expensive way.

Cowork is the easiest app to master. It can also 5x your token usage. Here's when to move up:

Step 1 · Cowork

You open it and delegate. A great place to start.

It's also the most expensive way to run Claude. You can't size the context window and there's no /compact. The window just grows, up to 1M tokens. That alone can 5x your token usage.

Cowork is a different machine than Claude Code: no hooks, no predefined subagents, no local MCP servers. Your project CLAUDE.md loads. Skills (and plugins) work, but global-install only.

When you start hitting usage limits or want to prototype or engineer with AI, move to step 2.

Step 2 · Claude Code in VS Code

It fixes the two things Cowork can't.

Cost control: you set the context window size and run /compact when you choose.

And the full kit: CLAUDE.md, rules, project skills, hooks, subagents, local MCP servers.

Most PMs can stop at this step. It's where I do most of my work.

Step 3 · The CLI (eventually)

The terminal version of the same Claude Code. Same kit: whatever you set up in step 2 is already there.

What it adds is scale and efficiency. Press the left arrow and every running session shows on one screen, jump into any of them. /cd switches the project folder mid-session.

How I split my work today:

→ Terminal: when I run multiple sessions in parallel
→ Extension: where I built my setup, and most of my knowledge work
→ Cowork: work that only needs Gmail, Drive, and a prompt

Tips for PMs:

→ In Cowork, use CLAUDE.md, not "project instructions."
→ Connect Gmail, Slack, and Drive once at claude.ai. They work in all 3 apps.
→ Global skills from Cowork are not visible in Code, and the other way round.
→ Skip global CLAUDE.md. It lives outside your repo, so GitHub sync can't reach it.
→ Skip learning "rules." You won't need them as a PM.
→ Start remote Claude sessions from your phone. My free template with a custom skill sets it up.

---

P.S. I recorded a walkthrough for each step. All free, templates included:

→ Cowork for PMs · start here, the easiest one
→ Claude Code for PMs · the full kit

All three free, one sign-up → https://lnkd.in/dqaRBzdu

How to set up Claude Code as a PM (no coding).

5 steps. Free template at the bottom.

1. Install the Claude Code extension in VS Code
→ VS Code → Extensions → search "Claude Code."
→ Install. Sign in with your Anthropic account.
→ Default to the extension, not the terminal.

2. Pick subscription or free
→ Pro ($20/mo) for most PM work.
→ Max ($100 or $200/mo) for bigger context.
→ Free: OpenRouter. The bonus template pre-wires it.
→ All three run on frontier models. They change weekly. Start anywhere comfortable.

3. Tell Claude about your product
→ Your strategy. Your product. Your value prop. Your team.
→ Ask Claude to save it to CLAUDE.md. It guides every action after that.
→ No filenames, no edits. Just describe what you do.

4. Connect your work apps
→ https://lnkd.in/dgeUfmFh → Add → pick the app.
→ Notion (PRDs), Linear or Jira (tickets), Slack (team), Gmail, Calendar, Drive.
→ Authenticate once. Every Claude Code session now sees the apps.

5. Install pm-skills
→ Type in chat: "install pm-skills."
→ 100+ PM skills and commands: discovery, PRDs, launches, OKRs, postmortems.
→ Plus segments, JTBD, competitor scans, persona builders.
→ Try first: "identify segments and jobs-to-be-done from these customer transcripts."

---

Bonus: everything in the free template
→ CLAUDE.md template (best practices baked-in)
→ Examples — custom skills, subagents, MCP server scaffold
→ Token reduction — tuned for 50–75% fewer tokens
→ Free-LLM settings — OpenRouter pre-wired

---

Anywhere you'd edit a file, ask Claude in the chat:

→ "Switch me to the free OpenRouter model."
→ "Update CLAUDE.md. We pivoted to enterprise."
→ "Create a skill for our launch postmortems."

You describe. Claude writes the file.

---

Start now. Have Claude Code running your next 1:1.

Free template & guide (no email, no paywall) → https://lnkd.in/dFnpuNwR

---

P.S. On Sep 1 I'm launching Claudathon for PMs. You learn by shipping production-ready apps. 200+ students so far. 4.8/5.0. No coding. Early bid: https://lnkd.in/dWNfApWB

4 types of RAG, and when to use which as a PM.

Three you embed in a product. One runs your agent:

1. Vanilla RAG
→ Embed docs, retrieve top chunks
→ For lookup questions
→ Demos and prototypes

2. Adaptive RAG
→ Cleans the query first
→ Picks a strategy per request
→ For messy, varied questions

3. Hybrid RAG
→ Queries several sources, merges
→ Vectors plus SQL plus search
→ For data spread across systems

4. Agentic RAG
→ The agent decides the retrieval
→ Files, tables, APIs, a vector store
→ Typically in your workspace agents

The first three are pipelines you design. The user asks → the system retrieves → the model answers. You decide the chunks, the sources, and what happens when the answer isn't in the data.

The fourth flips it. The agent decides what to read and where from: files, tables, an API, or a vector store. Mine run on markdown files. You'll mostly use agentic RAG in your workspace agents. For customer-facing products, careful orchestration and deterministic pipelines win most of the time.

Every week a PM asks me which vector store to pick:

1. If the docs fit in a prompt, paste them.
2. For ~1,000 docs → a folder of files.
3. For structured data, consider SQL.
4. Treat vector stores as the last resort.

The retrieval is rarely a challenge. It's the quality of your data. Missing documents, stale versions, PDFs that turned out to be scans.

---

P.S. Find this useful?

A free, interactive RAG simulator: https://lnkd.in/dsi6C5b9

I built the first two in n8n. No coding. Templates + step-by-step: https://lnkd.in/dew--RqD

AI agent architectures with n8n examples: https://lnkd.in/dnmQ_kP8

As a Claude subscriber you get a progress bar. And no idea where your tokens go. This gives you the full picture 👇

I created a Claude Code usage dashboard (token counts, costs, sessions, projects) directly inside VS Code. Reads local JSONL transcripts, no API calls.

Free, open source.

Find extension: Claude Code Usage by Paweł Huryn

Or install from VS Code marketplace with 1 click: https://lnkd.in/dKHfPdrT

Code on GitHub: https://lnkd.in/dN3P9pYj

Hope that helps.

P.S. The repo has 1.6K stars and 17 contributors. Now just shared as a VS Code extension, so you don't have to run scripts.

Most people use Claude the same way on day 100 as day 1. Same prompts. Same quality. Nothing compounds.

One block in your CLAUDE[.]md changes that.

I added this to every project. Claude now reviews what it learned before starting and extracts new insights after finishing — automatically.

Here's the full prompt:

------

Before starting a new task, review existing rules and hypotheses for this domain. Apply rules by default. Check if any hypothesis can be tested with today's work.

At the end of each task, extract insights. Store them in domain folders, e.g.:
‎ ‎ ‎ ‎ /knowledge/pricing/
‎ ‎ ‎ ‎ ‎ ‎ ‎ ‎ knowledge.md (facts and patterns)
‎ ‎ ‎ ‎ ‎ ‎ ‎ ‎ hypotheses.md (need more data)
‎ ‎ ‎ ‎ ‎ ‎ ‎ ‎ rules.md (confirmed — apply by default)

Maintain a /knowledge/INDEX.md that routes to each domain folder. Create the structure if it doesn't exist yet.

When a hypothesis gets confirmed 5+ times, promote it to a rule. When a rule gets contradicted by new data, demote it back to hypothesis.

------

One paste. Claude builds the structure, maintains it, and gets sharper every session.

------

To get all my takes without an algorithmic filter, subscribe to my newsletter: https://lnkd.in/dmCJgf4w

Product management is not about:
- Asking customers about the requirements
- Writing detailed specifications
- Mastering Scrum to perfection
- Obsessing over velocity, outputs, and the roadmap
- Acting like the CEO of the product

AI didn't make any of that good product management.

With AI, the pace changed dramatically:
- Summaries and status updates: minutes, not hours
- Understanding your data: a prompt, not a Jira ticket
- An early prototype: 15 minutes, or just ship the feature
- Iterating and releasing, several times a day

But the core of the job is still the same. It's about:
- Understanding the customers, the domain, and the business in depth
- Collaborating closely with engineers and designers
- Discovering problems worth solving
- Testing the riskiest assumptions
- Leading without authority
- Owning the outcome, especially when it's bad

One exception: AI agents joined the team. You lead them with context. But leading with context was always part of our job.

---

P.S. Find this useful? You can dive deeper, no paywall:

Product Discovery Guide, 2026 edition: https://lnkd.in/dWbnnjQX
WTF is an AI PM: https://lnkd.in/dBZUmDHq
Lead AI Agents Like Empowered Teams: https://lnkd.in/dhBzF-zB

GPT-6 Sol and Opus 5.5 just dropped. I don't trust the benchmarks, so I tested them on a real work: 2 repos, 105 hidden bugs, find and fix what you can. Blind judges.

Results:
- GPT-6 Astra (max): 45
- GPT-5.6 Sol (max): 43.5
- Fable 5.1 (max): 43
- Opus 5.5 (max): 41.7
- Muse Spark 1.3 (max): 32.2
- GPT-5.6 Luna (max): 31.3
- GPT-6 Sol (max): 29.3
- Opus 5 (max): 27
- GPT-6 Luna (max): 17 (one attempt, more in progress)

---

TLDR:

1. Opus 5.5 is a radical improvement, similar to Fable 5.1 for 67% of its cost. Use it by default for Claude. Free resources:

- How to get most of Opus 5.5 (the official guide): https://lnkd.in/d-DX8Xwj
- Claude Code for PMs: The Beginner's Guide: https://lnkd.in/dFnpuNwR

2. GPT-6 Sol looks like a radical degradation. But it’s also ~10x cheaper than GPT-5.6 Sol. Use as a budget option. Default to GPT-6 Astra or GPT-5.6 Sol, when working with Codex. Free resources:

- Codex App for PMs: https://lnkd.in/dqaRBzdu

3. GPT-5 Luna seems worse (-2) and more expansive (+40%) than DeepSeek V4.1 Flash (max). Might be a great option for simple, regular tasks, if you're worried about your data.

---

Note: All 105 are hard real bugs frontier model missed at the beginning of 2026. Don’t read it as “Astra solved only 45/105.”

Live scores, more models, and caveats (free, Bug Hunt Benchmark): https://lnkd.in/dYDq_9Qe