Skip to content

README.md

Latest commit

 

History

History
254 lines (154 loc) · 7.69 KB

README.md

File metadata and controls

254 lines (154 loc) · 7.69 KB

Example inputs gallery

Every pmstack capability with a runnable example input. Copy any block, paste into Claude Code (or Claude.ai with the skills installed), get a real artifact back.

If you don't have a real customer signal of your own yet, steal one of these to start — they're realistic, varied across domains, and produce non-trivial outputs.

Translation skills (signal → spec)

/prd — turn a signal into a PRD

/prd "Three of our biggest enterprise customers said code reviews are taking 24+ hours and devs are skipping them or merging without review. Two churned this quarter and named slow review as a top reason."

Variations to try:

  • A support-ticket signal:

    /prd "Support ticket #4421: 'Every time we onboard a new agent it takes 3 days for them to learn our routing rules. Half quit before they're productive.'"

  • An exec ask:

    /prd "CEO at Monday QBR: 'Why are we losing deals to Cursor on the AI-first IDE story? Our agent is better.' Build the PM response."

  • A churn quote:

    /prd "Churn interview, Acme Corp: 'We loved the product but our security team flagged 17 issues with how the agent stores conversation history. We didn't have time to wait for the fix.'"

Output: outputs/prd-<topic>-<date>.md (full 6-section PRD).

Reference: examples/walkthrough-code-review/prd-code-review-2026-05-05.md

/competitive — scan a market

/competitive "AI code review tools"

Variations:

  • /competitive "AI customer-support agents for e-commerce"
  • /competitive "developer-facing observability tools with AI features"
  • /competitive "low-code workflow builders that target operations teams"

Output: outputs/competitive-<market>-<date>.md (3–5 players, positioning, white-space).

Reference: examples/walkthrough-code-review/competitive-ai-code-review-2026-05-05.md

/compare — side-by-side two or more products

/compare "GitHub Copilot Code Review" "CodeRabbit"

Variations:

  • /compare "Cursor" "Windsurf" "Zed"
  • /compare "Claude" "GPT-4" "Gemini" "Llama" (for a model-pick decision)

Output: outputs/compare-<a>-vs-<b>-<date>-plan.md PLUS an executable eval YAML (outputs/compare-...-eval.yaml) you can run with /run-eval.

Reference: examples/walkthrough-compare-tools/

Spec-tightening skills

/metrics — design a measurement framework

/metrics "AI code review feature — measure whether it's actually speeding up reviews without missing bugs"

Variations:

  • /metrics "the recommendation engine on our product page"
  • /metrics "support agent triage speed"

Output: outputs/metrics-<topic>-<date>.md (North Star, 2–3 supporting, 1–2 counter-metrics, instrumentation plan).

Reference: examples/walkthrough-code-review/metrics-code-review-2026-05-06.md

/premortem — stress-test a draft PRD

/premortem prd-code-review

(The slug must match an existing PRD file in outputs/. Run /prd first if you don't have one.)

Variations:

  • /premortem prd-trial-onboarding
  • /premortem outputs/prd-support-routing-2026-04-25.md (full path)

Output: outputs/premortem-<topic>-<date>.md plus a confirmation gate to mutate the source PRD's Risks section.

Reference: examples/walkthrough-code-review/premortem-code-review-2026-05-05.md

Eval skills (the AI-PM-specific ones)

/eval — design a test suite for an AI feature

/eval "AI code review service — first-pass review with severity tags"

Variations:

  • /eval "support chatbot for an e-commerce site"
  • /eval "an agent that schedules customer meetings end-to-end"

Output: outputs/eval-<topic>-<date>.yaml (capabilities, failure modes, metrics, test cases).

Reference: outputs/eval-claude-ultraplan-2026-04-24.yaml and examples/walkthrough-code-review/eval-code-review-2026-05-06.yaml

/run-eval — actually run the eval

/run-eval outputs/eval-code-review-2026-05-06.yaml

Output: outputs/eval-runs/<run-id>/summary.md (pass-rates, top failures, metrics table, cost).

Reference: examples/walkthrough-code-review/eval-runs/code-review-eval-2026-05-06/summary.md

/eval-self — run pmstack against itself

/eval-self --skill prd

(Or /eval-self with no args to run the full suite. Costs ~$5–10 in tokens for the full run.)

Output: evals/results/<timestamp>_<model>.json plus a summary table.

Routine skills (the new five — schedulable)

/eval-drift — weekly drift watch

/eval-drift

Or schedule it: /loop 7d /eval-drift.

Output: outputs/eval-drift-<date>.md with RELEASE_BLOCKED: true|false flag.

Reference: examples/walkthrough-code-review/eval-drift-2026-05-12.md

/weekly — Monday self-snapshot

/weekly

Or schedule it: /loop 7d /weekly.

Output: outputs/weekly-<YYYY-Www>.md with three sections: Decisions made, Open loops aging, One thing I changed my mind about.

Reference: examples/walkthrough-code-review/weekly-2026-W19.md

/launch-readiness — pre-launch verifier

/launch-readiness code-review

Output: outputs/launch-readiness-<feature>-<date>.md with verdict (GO/NO-GO/CONDITIONAL) and 7-item evidence checklist.

Reference: examples/walkthrough-code-review/launch-readiness-code-review-2026-05-09.md

/lint — workspace audit

/lint

Or schedule it: /loop 7d /lint.

Output: outputs/lint-<date>.md with three sections: Graph gaps, Cross-artifact drift, Stale candidates.

Reference: examples/walkthrough-code-review/lint-2026-05-08.md

/premortem — see above (in spec-tightening section)

Stakeholder + orchestrator skills

/brief — stakeholder readout

/brief code-review exec

(Audiences: exec, engineering, customer, board, self.)

Output: outputs/brief-<topic>-<audience>-<date>.md.

Reference: examples/walkthrough-code-review/brief-code-review-exec-2026-05-09.md

/sprint — orchestrate prd → metrics → eval → brief

/sprint "Three of our biggest customers said code reviews are taking 24+ hours and devs are skipping them."

Output: four artifacts in sequence — PRD, metrics, eval YAML, brief — with confirmation gates between each.

/onboarding — interactive tutorial

/onboarding

Walks through every capability above using the bundled walkthrough. Recommended for first-time users.

How to use this gallery

  1. Pick the skill that matches the problem in front of you.
  2. Steal an example from this file — copy the input, replace one or two words to match your context.
  3. Run it. Read the output. Compare to the reference artifact.
  4. Iterate — refine the input, ask Claude follow-ups about the artifact, edit it.

If you produce a great artifact you'd be willing to share, open a PR adding it under examples/.