Organize Claude Code skills as one folder per skill under .claude/skills/, each with a SKILL.md file, grouped mentally by PM workflow stage: discovery, strategy, delivery. The ceiling on collection size is routing, not count. A hand-curated setup works at 5 to 10 skills because you are the router; a routed system like AI PM OS works at 243 because guided workflows and context indexing invoke the right skill for you.
A skill is a folder with a SKILL.md inside: YAML frontmatter carrying a name and a description, then markdown instructions. Claude Code reads the descriptions of every skill in your project and invokes one automatically when your request matches. You rarely call a skill by name.
That routing mechanic drives everything else on this page. The description is the API: a precise "Use when... Not for..." description gets the skill invoked at the right moments, and two skills with overlapping descriptions will steal each other's invocations. This is why an unrouted pile degrades as it grows, and why routing is the real skill-management problem for a product manager in 2026.
One folder per skill, each containing its SKILL.md. Inside a PM workspace it looks like this:
product-hub/
├── AGENTS.md # product context, conventions, hard rules
├── CLAUDE.md # one line: @AGENTS.md
├── decisions.md # decision log
├── research/ # transcripts, notes, synthesis
├── prds/ # dated PRD versions
└── .claude/
└── skills/
├── interview-synthesis/
│ └── SKILL.md
├── decision-log/
│ └── SKILL.md
├── prd-review/
│ └── SKILL.md
└── stakeholder-update/
└── SKILL.mdTwo placement rules. Skills in the project's .claude/skills/ travel with the repo, so your team gets them via git. Personal skills you want in every project go in ~/.claude/skills/ instead. Supporting files (templates, examples, checklists) live inside the skill's own folder next to its SKILL.md.
The four skills in the tree above are real and free: the AGENTS.md for product managers page ships complete, paste-ready SKILL.md files for each, plus the AGENTS.md context file they assume.
Group by the question each stage answers, not by artifact type. Folder names stay flat under .claude/skills/; this grouping is how you plan, audit, and spot gaps in the collection.
What should we build?
Skills that turn raw user signal into evidence. The anti-invention rule matters most here: every claim must trace to a quote or a data point.
interview-synthesis, survey-analysis, jtbd-extraction, competitor-teardown
Why this, why now?
Skills that force options and trade-offs onto paper. A decision-log skill pays for itself the first time a decision gets re-litigated.
problem-framing, prioritization, decision-log, bet-memo
Is it shipping well?
Skills that catch gaps before engineering does and keep stakeholders informed without a meeting.
prd-review, gherkin-stories, release-notes, stakeholder-update
The enemy is not a big collection; it is an unrouted pile. Invocation runs on descriptions, so a DIY folder of a hundred overlapping descriptions means the wrong skill fires or none does. Unaudited skills drift out of date and inject stale instructions into real work. And when a skill misfires during a deadline, you stop trusting skills, and the whole setup quietly dies.
If you curate by hand, you are the router, so keep the collection small: one skill per task you repeat weekly. For most PMs that is a synthesis skill, a decision-log skill, a PRD review skill, a stakeholder update, and two or three specific to your product area. Add a skill only after you catch yourself typing the same multi-step instructions twice. Prune quarterly.
A routed system removes that ceiling. This is exactly what AI PM OS is: 243 PM skills that stay out of each other's way because they are indexed by context and invoked through 11 guided workflows, so the routing that limits a hand-built pile is done by the system instead of by you. You can browse the full AI PM OS skill catalog by category to see how a curated collection is organized in practice.
Four complete PM skills (interview synthesis, decision log, PRD review, stakeholder update) plus a full AGENTS.md, ungated, on the AGENTS.md template page.
deanpeters/Product-Manager-Skills (70+ skills) and phuryn/pm-skills offer breadth. Quality varies by skill, so read each SKILL.md before installing, and keep only what maps to your weekly loop.
AI PM OS ships 243 skills, 11 guided workflows, and 12 subagents for Claude Code and Codex, with the curation and routing already done.
A skill is a folder containing a SKILL.md file: YAML frontmatter (a name plus a description of when to use it) followed by markdown instructions. Claude Code scans .claude/skills/ in your project, reads the descriptions, and invokes a skill automatically when your request matches. You do not call skills by name; the description does the routing, which is why a precise "Use when... Not for..." description matters more than the instructions themselves.
One folder per skill under .claude/skills/, each containing a SKILL.md, for example .claude/skills/interview-synthesis/SKILL.md. Project-level skills live in the repo (shareable with your team via git); personal skills can live in ~/.claude/skills/ and apply everywhere. Keep supporting files (templates, examples) inside the skill folder next to its SKILL.md.
It depends on who does the routing. If you curate a folder by hand, you are the router, so keep it to 5 to 10 skills covering your weekly loop: an unrouted pile with overlapping descriptions means the wrong skill fires or none does. Routed systems scale past that ceiling, which is how AI PM OS runs 243 skills: guided workflows and context indexing do the routing instead of you. For a DIY setup, start with one skill per recurring weekly task, and add a new one only when you catch yourself giving the same multi-step instructions twice.
By workflow stage: discovery (what should we build), strategy (why this, why now), and delivery (is it shipping well). Artifact-based organization (prd/, roadmap/, notes/) sounds tidy but hides the question each skill answers, and questions are how you actually reach for skills. The folder names themselves stay flat under .claude/skills/; the stage grouping lives in how you plan and audit the collection.
Three sources. Free samples: the AGENTS.md page on this site ships four complete PM skills, ungated. Free breadth packs: GitHub collections like deanpeters/Product-Manager-Skills (70+ skills) and phuryn/pm-skills offer wide coverage, quality varies by skill, so audit before installing. Curated system: AI PM OS ships 243 PM skills with 11 guided workflows and 12 subagents, indexed so the right skill fires by context instead of you managing the collection.
Three properties. First, a description that says when to use it and when not to, because routing runs on descriptions. Second, steps that each end in a checkable completion criterion ("done when every claim cites a quote"), so the agent cannot skip ahead. Third, hard rules against invention: a synthesis skill that lets the model fabricate user quotes is worse than no skill at all.
AI PM OS is 243 PM skills organized so the right one fires by context, plus 11 guided workflows and 12 subagents. Built for Claude Code and Codex.
Written by George Nurijanian, Product Manager with 8+ years of experience. For the complete system, see AI PM OS.