10x-pm

Prompts don't compound. Skills do.

The open-source AI toolkit for product managers — 13 skills and 3 red-team agents for Claude Code and Claude Desktop, covering the full PM job from napkin idea to shipped launch.

Validate License: MIT PRs Welcome

Install · What's inside · See it work · Philosophy · FAQ

Every PM has a folder of prompts. It doesn't change how they work, because a prompt is advice — the model can take it or leave it, and next week you paste it again from scratch.

A skill is different. It's installed judgment: the questions a good PM asks before writing, the structure reviewers actually want, and a quality bar the AI must self-check against before it shows you anything. Install once; every PRD, prioritization call, and exec update starts from that bar instead of a blank page.

Built by a practicing PM running an AI Center of Excellence. Everything here is generalized from workflows that run on real product work — not template cosplay.

Install (60 seconds)

Claude Code:

claude plugin marketplace add ramanbamba/10x-pm
claude plugin install 10x-pm@10x-pm-marketplace

Then just talk: "write a PRD for bulk export", "triage this feedback CSV", "pre-mortem my launch plan". Skills trigger automatically.

Claude Desktop / Cowork: grab 10x-pm.plugin from Releases and open it. Done.

No plugin support? Copy any skills/<name>/ folder into your project's .claude/skills/ directory. Each skill is self-contained.

What's inside

13 skills — the PM workflow

Skill What it refuses to let you ship Trigger
prd Problem statements with solutions hiding inside; metrics without numbers and dates "write a PRD for…"
prioritize Rankings where nothing hurts; 100% confidence on guesses "help me prioritize…"
user-interviews Themes built on one participant; interpretation dressed as observation "synthesize these interviews"
feedback-triage "Many users" when the number is 7; volume mistaken for pain "triage this feedback"
competitive-teardown Feature tables as conclusions; undated intel; teardowns that find nothing to admire "tear down [competitor]"
metrics-tree North stars that count activity, not value; drivers without guardrails "define metrics for…"
exec-update Buried bad news; asks without deadlines; updates over 250 words "write my exec update"
story-splitter Horizontal layer-stories; splits where nothing ships until story 5 "split this epic"
roadmap Bets without kill criteria; fake dates; empty "not doing" lists "build a roadmap"
launch-plan Rollback triggers that are moods, not numbers; unowned workstreams "plan this launch"
experiment-design Unfalsifiable hypotheses; underpowered tests recommended silently "design an experiment"
stakeholder-comms Buried "no"s; fake-open decisions; blame laundering "help me tell [X] that…"
release-notes "Various bug fixes and improvements" "write release notes"

3 agents — the red-team layer

Run these against your own documents before anyone else sees them. This is the part prompt libraries can't do: adversarial review with a fresh context, no author's blind spots.

Agent What it does
customer-voice Reads your PRD as your actual target user — busy, skeptical, carrying existing tools — and reports the exact line where they'd stop reading
pre-mortem Assumes your launch already failed and works backwards across five failure lanes (demand, execution, distribution, organization, external) to the kill shot
exec-reviewer Reviews your memo the way a sharp VP does — 90-second skim first — then lists the five questions you'll get ambushed with in the room 
> Use the pre-mortem agent on docs/launch-plan.md

Bonus: CLAUDE.md for product teams

templates/CLAUDE.md — a drop-in context file that makes Claude behave like a disciplined product thinker in any repo or docs workspace: challenge assumptions, demand evidence, keep documents short. Steal it even if you install nothing else.

See it work

Real interaction, sanitized (full version + more in examples/):

You: write my exec update — we integrated the payments API, hit some SDK issues so checkout slips maybe 2 weeks, NPS survey went out, hired a designer

Without the skill: 400 words starting with the wins, the slip in paragraph four, no ask.

With exec-update: asks you two questions (original commitment date? what do you need from readers?), then:

Checkout — July 4

Headline: Checkout slips ~2 weeks (Aug 15 → Aug 29) on payment-SDK defects; need your OK on the revised date by Friday or we descope saved-cards from v1.

Status: 🟡 At risk → vs. commitment of Aug 15 …(198 more words, bad news first, one decidable ask)

The difference isn't writing quality. It's that the skill refused to write until it knew the original commitment — because status against original commitment is what makes an update honest.

Why skills, not prompts

  1. Interrogate before generating. Every skill asks the 2–4 questions that change the output before writing a word. A PRD built on unexamined assumptions is fast garbage.
  2. A quality bar, not just a template. Each skill ends with explicit failure modes the AI must self-check: vague metrics, unfalsifiable hypotheses, roadmaps with no kill criteria, "various bug fixes."
  3. Evidence over vibes. Synthesis skills separate observation from interpretation and carry verbatim quotes forward. Counts are computed, never estimated.
  4. Short beats complete. Exec updates cap at 250 words. PRDs cap at two pages. Nobody reads the 9-page PRD, and the skills know it.
  5. Adversarial review is built in. Write with a skill, then attack the output with an agent. The loop — draft → red-team → revise — is the actual 10x, not the drafting.

FAQ

Do I need to be technical? No. If you can install Claude Code (one command) or Claude Desktop (an app), you can use all of this. The PM's setup guide takes you from zero in 15 minutes.

Does this work with ChatGPT/Gemini/Copilot? The SKILL.md files are markdown — the judgment inside is portable anywhere. The auto-triggering and agents are Claude-specific.

Will this leak my company data? The toolkit is local files with no telemetry, no external calls, no accounts. Your data goes wherever your Claude setup already sends it, and nowhere else.

How is this different from a prompt library? Prompts are advice. Skills are enforced process: they interrogate first, self-check against failure modes, and trigger automatically. See Philosophy.

Can I adapt these to my company's templates? Yes — that's the point. Fork it, edit the output formats to match your org, keep the quality bars. MIT license.

Roadmap

  • pricing-analysis — packaging and price-change skills
  • board-deck — quarterly narrative builder
  • discovery-interview — live interview guide generator (not just synthesis)
  • Connector-aware variants (Jira, Linear, Productboard via MCP)
  • Eval suite — regression tests for skill output quality

Vote with issues: propose a skill.

Contributing

The bar for PRs: encode judgment, not templates. The most valuable contribution is a new failure mode for an existing skill's quality bar — a way outputs go wrong that the skill should catch. See CONTRIBUTING.md.

License

MIT — use it, fork it, bring it to your team.

If this saved you an hour, star the repo — it's how other PMs find it.

Maintained by Raman Bamba. Weekly breakdowns of AI-native PM workflows on LinkedIn.