# Claude for AI Product Managers: 36 Claude Skills to Test an AI Feature Before It Ships > From "do we even need AI" to the weekly transcript read — 36 skills that spec an AI feature as tests, build its evals from real outputs and leave the verdict to a named person. 36 free Claude Skills (Agent Skills) by Pauline Bertry, Polar Bear. For Claude Code, Claude.ai and Claude Desktop. - Page: https://meet-polar-bear.com/skills/ai-product-managers-pack - Download: https://meet-polar-bear.com/skills/ai-product-managers-pack.zip - Source: https://github.com/polar-bear-org/claude-skills/tree/main/packs/ai-product-managers-pack - License: Free to use inside your company, not for resale. - For: AI product managers, product managers adding AI features, founders shipping AI products ## Install **Claude Code** — two lines, every skill in the pack at once: ``` /plugin marketplace add polar-bear-org/claude-skills /plugin install ai-product-managers-pack@polar-bear-skills ``` **Claude.ai / Claude Desktop** — download https://meet-polar-bear.com/skills/ai-product-managers-pack.zip and upload each skill from its `install/` folder under Settings → Customize → Skills. **Manually** — copy `packs/ai-product-managers-pack/skills/*` from https://github.com/polar-bear-org/claude-skills into `~/.claude/skills/` or a project's `.claude/skills/`. ## The 36 skills ### 1. Aipm AI use case canvas (`aipm-ai-use-case-canvas`) Canvas (prediction, judgment, action, outcome, input, feedback), simpler-fix check, data check, verdict Say: "run aipm-ai-use-case-canvas", "AI canvas", "do we even need AI for this", "leadership wants an AI feature", "is this a good AI use case" ### 2. Aipm AI failure modes (`aipm-ai-failure-modes`) Failure list per output, false yes vs false no cost, severity and detection, graceful failure, tolerable rate left blank Say: "run aipm-ai-failure-modes", "AI FMEA", "which AI mistakes matter", "leaders expect it to be right every time", "false positives vs false negatives" ### 3. Aipm workflow or agent (`aipm-workflow-or-agent`) Candidate patterns, cost, latency and failure trade-offs, simplest pattern that passes, trigger for the next step up Say: "run aipm-workflow-or-agent", "do we need an agent", "workflow vs agent", "agent or prompt chain", "agentic architecture choice" ### 4. Aipm AI model selection (`aipm-ai-model-selection`) Side by side test plan on real inputs, quality, latency and cost table per tier, pick per task, re-check date Say: "run aipm-ai-model-selection", "which model tier should we use", "justify the model choice", "compare models side by side", "cheaper model or stronger model" ### 5. Aipm AI prototype brief (`aipm-ai-prototype-brief`) What the demo tests, pass line set in advance, cherry-picked input check, gap-to-production list Say: "run aipm-ai-prototype-brief", "leadership thinks the demo is the product", "AI demo to production", "test card for an AI prototype", "what does this AI demo prove" ### 6. Aipm AI prd (`aipm-ai-prd`) Problem, inputs the model sees, outputs, behavior summary, fallbacks, data needs, release bar, open questions Say: "run aipm-ai-prd", "write a PRD for an AI feature", "AI feature spec", "spec for our copilot", "PRD template for an LLM feature" ### 7. Aipm AI success criteria (`aipm-ai-success-criteria`) Measurable criteria per dimension, today's baseline, target set by a named person, how each is measured Say: "run aipm-ai-success-criteria", "define success criteria for our AI feature", "what does good mean for this model", "make it good is not a spec", "acceptance criteria for an LLM" ### 8. Aipm behavior contract (`aipm-behavior-contract`) Must, must never and when-unsure lines as Given-When-Then cases, pass rate over repeated runs, guardrails Say: "run aipm-behavior-contract", "behavior spec for our AI", "must never rules for the assistant", "Given When Then for an LLM", "turn helpful and safe into tests" ### 9. Aipm human in the loop (`aipm-human-in-the-loop`) Approve, edit or take-over points, triggers, handoff message, response time a person can meet, log Say: "run aipm-human-in-the-loop", "where should a human approve", "human in the loop design", "AI handoff to a person", "approval step for the agent" ### 10. Aipm data privacy brief (`aipm-data-privacy-brief`) Data flow from user to model to logs, personal data in prompts, retention, region, questions for legal Say: "run aipm-data-privacy-brief", "what user data reaches the model", "data flow for our AI feature", "privacy review for an LLM feature", "legal asked about the model's data" ### 11. Aipm system prompt (`aipm-system-prompt`) Role, audience, task, rules, examples, output format, refusals and handoffs, critique of the current prompt Say: "run aipm-system-prompt", "write the system prompt", "clean up our prompt", "system prompt template", "which line of the prompt does what" ### 12. Aipm context engineering (`aipm-context-engineering`) Knowledge sources and owners, freshness rules, what to leave out, retrieval spot checks, context budget Say: "run aipm-context-engineering", "context engineering", "which documents should the model read", "the AI answers from stale docs", "knowledge base for our AI feature" ### 13. Aipm agent spec (`aipm-agent-spec`) Goal, allowed tools, permissions, stop conditions, spend budget, approval points, never-alone list Say: "run aipm-agent-spec", "agent spec", "write the limits for our agent", "agent permissions", "where should the agent stop" ### 14. Aipm tool descriptions (`aipm-tool-descriptions`) Per tool name, purpose, inputs, outputs, error messages, namespacing, overlap check, three test tasks Say: "run aipm-tool-descriptions", "write tool descriptions", "the agent picks the wrong tool", "tool definitions for our agent", "MCP tool descriptions" ### 15. Aipm error analysis (`aipm-error-analysis`) Notes on real outputs, open codes, failure types grouped, counts per type, the three to fix first Say: "run aipm-error-analysis", "error analysis", "what is actually going wrong with our AI", "read our transcripts", "build a failure taxonomy" ### 16. Aipm golden dataset (`aipm-golden-dataset`) 20 to 50 real cases with expected outcomes, coverage by failure type, source and version, add-and-retire rules Say: "run aipm-golden-dataset", "golden dataset", "build an eval set", "test cases for our AI feature", "did the new prompt actually help" ### 17. Aipm synthetic test data (`aipm-synthetic-test-data`) Generated cases for coverage gaps only, each marked synthetic, checked by a person, kept out of the headline score Say: "run aipm-synthetic-test-data", "synthetic test data", "generate edge cases", "we do not have enough test cases", "test cases for rare inputs" ### 18. Aipm eval rubric (`aipm-eval-rubric`) One pass or fail check per failure type, grader per check, pass and fail examples, blockers Say: "run aipm-eval-rubric", "eval rubric", "pass fail criteria for our AI", "how should we grade outputs", "our 1 to 5 scores mean nothing" ### 19. Aipm llm judge (`aipm-llm-judge`) Judge prompt for one failure type, labelled set, agreement on a held-out split, bias checks Say: "run aipm-llm-judge", "LLM as a judge", "judge prompt", "automate our eval grading", "can we trust the LLM grader" ### 20. Aipm eval plan (`aipm-eval-plan`) Capability and regression suites, when each runs, what blocks a release, named owner of quality Say: "run aipm-eval-plan", "eval plan", "AI eval strategy", "who owns quality", "when should evals run" ### 21. Aipm red team plan (`aipm-red-team-plan`) Attack cases by risk class, who runs them, pass line, fixes before launch Say: "run aipm-red-team-plan", "red team our AI feature", "prompt injection test plan", "the agent reads emails from strangers", "OWASP LLM top 10 checklist" ### 22. Aipm AI ux review (`aipm-ai-ux-review`) 18 interaction guidelines checked by phase, AI disclosure check, correction and handoff paths, fixes ranked Say: "run aipm-ai-ux-review", "AI UX review", "human-AI interaction guidelines", "users keep rephrasing", "what happens when the assistant is wrong" ### 23. Aipm AI impact assessment (`aipm-ai-impact-assessment`) Who is affected, intended use and misuse, data, oversight, mitigations, residual risk owner, adviser questions Say: "run aipm-ai-impact-assessment", "AI impact assessment", "algorithmic impact assessment", "the customer wants an impact assessment", "foreseeable misuse of our AI feature" ### 24. Aipm AI risk register (`aipm-ai-risk-register`) Cause, event, effect risks mapped to Map, Measure, Manage, owner, trigger, response Say: "run aipm-ai-risk-register", "AI risk register", "NIST AI RMF risks", "legal and security want our AI risks", "turn these worries into risks" ### 25. Aipm legal questions (`aipm-legal-questions`) Questions for the legal and privacy team by topic, with the facts attached; asks, never answers Say: "run aipm-legal-questions", "questions for legal", "prepare for the legal review", "what do I ask legal about our AI feature", "legal sign-off for AI" ### 26. Aipm AI launch checklist (`aipm-ai-launch-checklist`) Go-live list with an owner per line, from evals passed as run to rollback rehearsed Say: "run aipm-ai-launch-checklist", "AI launch checklist", "go-live checklist for an AI feature", "are we ready to launch", "launch readiness review" ### 27. Aipm launch gates (`aipm-launch-gates`) Stages, eval and live numbers per gate, kill criteria, rollback path, who signs each gate Say: "run aipm-launch-gates", "AI launch gates", "staged rollout for an AI feature", "canary release plan", "kill criteria for our AI launch" ### 28. Aipm quality review (`aipm-quality-review`) Transcript sample plan, review notes, new failure types, cases added to the golden dataset, decisions Say: "run aipm-quality-review", "weekly AI quality review", "read our AI transcripts", "it worked for weeks then broke", "review production conversations" ### 29. Aipm feedback signals (`aipm-feedback-signals`) Signal map (accept, edit, retry, rephrase, abandon, escalate, rating), logging, weekly readout, review triggers Say: "run aipm-feedback-signals", "users never rate our AI answers", "implicit feedback for an AI feature", "what should we log for our assistant", "thumbs up thumbs down is not enough" ### 30. Aipm prompt regression test (`aipm-prompt-regression-test`) Change note, cases to re-run, before and after per failure type, ship or hold for a named person Say: "run aipm-prompt-regression-test", "did this prompt change break anything", "regression test a prompt edit", "before and after eval for our prompt", "the fix broke another case" ### 31. Aipm model migration (`aipm-model-migration`) Deadline, breaking changes, eval rerun, cost difference, behavior diffs to read, rollout and fallback Say: "run aipm-model-migration", "our model is being retired", "deprecation notice for our model", "migrate to the replacement model", "model upgrade plan" ### 32. Aipm AI incident response (`aipm-ai-incident-response`) Severity levels, containment, customer correction, evidence timeline, fix into the golden dataset, blameless review Say: "run aipm-ai-incident-response", "our AI told a customer something wrong", "AI incident playbook", "the chatbot made a promise we cannot keep", "switch off the AI feature" ### 33. Aipm unit economics (`aipm-unit-economics`) Cost per call, per task and per successful outcome, heavy-user case, caching and batch levers, today's cost Say: "run aipm-unit-economics", "cost per task", "cost per successful outcome", "what does each AI answer cost us", "finance wants to stop the feature" ### 34. Aipm usage and pricing test (`aipm-usage-and-pricing-test`) Small real-task test, usage estimate at light, normal and heavy use, price options to test, what to re-measure Say: "run aipm-usage-and-pricing-test", "what will this feature cost to run", "estimate AI usage before launch", "seat or usage pricing for AI", "credits pricing test" ### 35. Aipm feature card (`aipm-feature-card`) Intended and out-of-scope use, data, eval results as run with dates, known limits, disclosures, owner Say: "run aipm-feature-card", "model card for our feature", "what does the AI feature do and where does it fail", "AI fact sheet for sales", "document the AI feature for legal" ### 36. Aipm exec brief (`aipm-exec-brief`) One page, bottom line first: how often it is wrong by severity, what that costs, run cost, the decision asked Say: "run aipm-exec-brief", "how accurate is it", "explain the AI to leadership", "one-pager for execs on the AI feature", "go or no-go memo for the AI"