Agent PM: AI Agent Product Review and Fixing
Paste the following prompt into your AI chat to install this skill:
Please install @user_88546431/agent-pm-user-88546431 according to https://skillhub.cn/install/skillhub.md.
About this skill
Problem It Solves
Many AI Agent projects fail not because code cannot run, but because instructions, tool descriptions, workflows, and goals are under-specified: skill.md is vague, tools lack fallbacks, output formats conflict, and users must guess behavior. Agent PM treats these as product-design issues. Its goal is not to produce an endless issue list, but to create a verifiable, bounded review-fix loop.
How It Works and Limits
It works through four modes:
- Review: checks intent clarity, usability, robustness, minimum complexity, and consistency; candidate issues must pass context reading, execution simulation, layer separation, concept-vs-implementation distinction, and logic-chain validation.
- Fix: focuses on confirmed P0/P1 issues, requires acceptance criteria before patching, and checks for regressions caused by the fix.
- Product scoring: can be entered independently to assess goals, workflow, output, and user experience without making code changes.
- New Skill clarification: asks about goals, boundaries, triggers, inputs, outputs, and resource dependencies before file creation.
It applies to Skill files, CLAUDE.md, instruction files, and MCP Server tool design. It is not general code review, does not cover non-Claude Code ecosystems such as LangChain or AutoGen, and does not block on P2/P3 issues by default.
Use Cases
- Review a Claude Code skill file to check intent, tool calls, and output format conflicts.
- Inspect MCP Server tool design for missing fallbacks, retries, or responsibility conflicts.
- Clarify a new Skill before generation to close goals, triggers, I/O, resources, and acceptance criteria.
- Score the current project state after review and decide whether P0/P1 fixes must block scoring.
Best For
- Engineers maintaining Claude Code skills who need to find unclear intent, boundaries, or outputs in skill.md.
- Product engineers owning MCP toolsets who need to check tool descriptions, fallbacks, and responsibility stability.
- Owners preparing a new Agent skill who need to close goals, triggers, resources, and acceptance criteria first.
- Product managers accepting Agent projects who need to push fixes based on P0/P1 blockers and score conclusions.
Related Skills
Automatically searches job postings based on the user profile, AI-scores fit, saves desktop reports, and sends email updates with scheduled tracking.
Activates bionic reasoning for causal judgment, numerical prediction, and hypothesis validation, using hypothesis-driven checks, Bayesian updates, falsifiability tests, bias defense, and physical constraints.
Triggered by /plan, it asks the agent to output a plan, risks, impact scope, and validation approach before execution.
Track and clean agent session files, packages, and Skills via trash-first safety.