🧠 PM Skills: Professional Agent Skills for Claude, ChatGPT, Gemini, Cursor, Codex & Hermes
One request to your AI, one finished deliverable back: the PRD, the stakeholder update and the decision call, at a senior product manager's standard.
Try it right now. Install below, then paste this into your assistant:
Here are my discovery notes from six customer calls: [paste your notes].
Turn them into a one-page PRD, a risk register with owners, and a
stakeholder update I can send to leadership today.
Your assistant loads user-interview-synthesis, and and returns the three documents, not advice about writing them. Not a PM? The same library covers engineers, job seekers, students, founders and 中文用户, below.
Key features
01. 1310 professional Agent Skills + workflow recipes — searchable & fetchable over MCP
Your landlord kept your deposit. You got laid off on a Tuesday. Your boss wants the PRD by Friday.
Generic AI is a very confident intern. PM Skills is the senior colleague's notes: 1320 of them, one markdown file each. (PM stands for Professional. Yes, we get asked.)
PM Skills is an open-source library of 1320 Agent Skills — plain-markdown SKILL.md files that teach an AI assistant to do one professional task to a senior professional's standard, from writing a PRD to decoding a lease or running a blameless postmortem. Each skill bundles the framework, an output template, quality checks, and anti-patterns. It is MIT-licensed and works with Claude, ChatGPT, Gemini, Cursor, and Codex.
⚡ Install in 30 seconds
npx pm-claude-skills add # any tool: Cursor, Codex, Windsurf, Cline, Zed…
# In Claude Code: /plugin → search pm-skills (official Anthropic directory)
npx --registry=https://registry.npmmirror.com pm-claude-skills add # 中国大陆镜像
Then just say what you need: "Decode this job ad and tell me where I am weak." Nothing here can break your setup: a skill is a markdown file your AI reads; installing copies text files, uninstalling deletes them. Read one first. Per-tool instructions: docs/installation.md · ChatGPT and Gemini: copy from exports/ · MCP: claude mcp add pm-skills -- npx -y -p pm-claude-skills pm-claude-skills-mcp · No install at all: point any MCP client at the hosted broker https://pm-skills-mcp.pm-claude-skills.workers.dev/mcp (how it works) · Starting a team? clone a starter template for your role · No network? every release ships an offline bundle you can serve on a laptop
🔄 Keep it updated
Skills improve every week. Whichever way you installed, updating is one step.
You installed with
Update with
npx pm-claude-skills add (any tool)
Run the same command again. npx pm-claude-skills doctor lists what is installed and what is stale.
The Claude Code marketplace
Open /plugin, pick the pm-claude-skills marketplace and choose Update.
The Python package
pip install -U pm-skills
The MCP server
Use npx -y -p pm-claude-skills@latest pm-claude-skills-mcp in your MCP config and every start is the latest release.
A fork of this repository
The bundled sync-fork.yml workflow fast-forwards your main every week. See docs/fork-sync.md.
🌅 Claude Duet, OpenAI Dots and other proactive agents
PM Skills works with Claude Duet, Anthropic's proactive agent. Duet uses the skills in your Claude account, so every skill here can run inside it. Agents that work while you're away fail in new ways: briefs that pad, uncited claims, acting instead of drafting, saying "done" about a draft. So 21 Duet-ready skills are built for unattended use. They keep briefs short and cited, draft instead of sending, and always ask before anything is sent, posted, paid, signed or deleted.
The five skills: prepare-for-duet · standing-orders · approval-policy-designer · morning-brief-designer and interrupt-rules · agent-kill-switch-plan
Into Claude and Duet: press ⬇ Add to Claude on any skill's page, then upload the ZIP in claude.ai (Settings, Skills, Upload skill).
A starter pack of 3 or 4 skills (founder, freelancer, shop owner, team lead, business ops) on the Duet page, or npx pm-claude-skills add --agent claude --skills standing-orders,cash-watch,store-watch.
The whole bundle:npx pm-claude-skills add --bundle pm-duet.
Test it safely: the injection drill has 12 planted instructions to check whether your agent reports or obeys them. The fake-apps sandbox gives invented Gmail, Stripe and Shopify data where every send or payment is refused and logged.
Honest status: every one of these skills is designed for Duet. None has been tested on Duet yet, because it's a waitlisted research preview. File a Tested on Duet report and a skill moves up a tier, with your name on it. They're written to work with Dots and Muse too: see how the three compare.
🥊 The demo
More to see: the Showcase gallery has the Playground, the tech tree, Galaxy 3D and Skill City.
🧭 How it works
A skill is one plain-markdown SKILL.md file: a description that tells your assistant when to load it, plus the working framework, output template and quality checks it follows. The installer (npx pm-claude-skills add) copies the skills you choose into the config folder of the tool you use, so they load only when a request matches. The MCP server serves skills on demand to any MCP client instead, and the same package exposes a read-only REST API for n8n, Make and other automations. No runtime, no account, nothing resident: uninstalling deletes text files.
🧑🤝🧑 Pick your path
The product-management core first, then the bundles everyone borrows. Each one is a real bundle in plugins/.
Install any bundle: /plugin install pm-decoders@pm-claude-skills, or npx pm-claude-skills add --bundle pm-decoders. All 154 bundles with every skill and install command: docs/catalogue.md · every skill in one page: SKILLS.md · chain them: WORKFLOWS.md
21 Duet-ready skills for Claude Duet and other proactive agents: standing orders, approval policies, a kill switch, small-business watchers (cash, payroll, signatures, CRM, store), team charters, an injection drill, a fake-apps sandbox, starter packs and an Add to Claude ZIP for every skill
"Write standing orders for my Duet" · "Watch my Stripe and Xero every morning"
Easy Read, Teach me, voice and mixed-language questions, local rules for your place, a skill for every form, newcomer, informal-economy, disability and after-a-disaster skills, the library offline on a Raspberry Pi, a USSD menu for any phone, right-click any file, skill apps, partner checks and an expert-graded jobs exam
"Explain this P60 to me" · "मेरा landlord deposit नहीं दे रहा"
v82.0.0 China for everyone, Korea and Southeast Asia
Teachers, parents, trade and manufacturing packs, 适老化, Korean, Japanese and Southeast Asian packs, lite skills for local models, Word and WPS add-ins, team sync
An honest summary of what is checked, measured and still unreviewed, computed from the repo on every change:
Checked on every change, in CI: structure (SkillCheck, all 1,320 skills at the L3 gate), a security audit, a near-duplicate check, a count-drift check, translation parity (178 translations in 10 languages), and a prompt-injection suite of 15 hostile documents. A change that fails any of them does not merge.
Eval cases: 424 of the 1,320 skills carry eval cases (evals/cases.json), and 163 have a worked example in examples/.
Measured scores: 399 skills have blind-judged scores on the leaderboard (average 4.6/5 across the models tested); the benchmark report publishes the negative findings too. Everything unscored is unscored, not assumed good.
Checked by people who do the job: organisations can vouch for skills as partners, and the jobs exam has AI answers marked blind by working professionals on their own marking schemes. Both start empty and fill only with real participants.
It rewrites itself, in public: the weakest skills are periodically rewritten by a model and kept only if a judge scores the rewrite higher; every attempt, including the losses, is in the evolution log.
Honesty rules baked in: decoders end with a not-advice line (the severity scale), calculators name what they do not model, simulators debrief out of character. PM Skills is not a substitute for a professional.
Privacy, stated plainly: everything installable runs locally. Skills are inert markdown on your machine; the CLI, the exports and the local MCP server send nothing anywhere, and telemetry is off unless you set PM_SKILLS_TELEMETRY=1 (skill name only, never content). The two things that do leave your machine are your prompts to whichever model provider you chose, and requests to the hosted playground or hosted MCP if you opt to use them. Full statements: docs/TELEMETRY.md · report a vulnerability: SECURITY.md · licence: MIT. Verify a release: npx pm-claude-skills verify --release (signed SBOM, Sigstore provenance). Verify one skill: its SHA-256 is published in /.well-known/agent-skills.json, so running shasum -a 256 on your copy of the file should match; every exported file carries the same hash in a provenance footer. Agent-native skills are red-teamed on every change (evals/adversarial).
🙌 Community
Show and Tell: share what a skill produced, including what it got wrong: open a Show and Tell. The best entries become showcases.
Vote on what gets built next: the tech tree ranks open skill requests by 👍; the top three are in research.
Credits follow the all-contributors spec (.all-contributorsrc). Contributed and not listed? Open a pull request running node scripts/build-contributors.mjs --add <your-login> <type>, or ask in an issue. Translators: pick a skill from the claim board.
