liqiongyu

Create an AI Evals Pack (eval PRD, test set, rubric, judge plan, results + iteration loop). See also: building-with-llms (build), ai-product-strategy (strategy).

Create an AI Product Strategy Pack (thesis, use cases, system plan, eval plan, roadmap).

Analyze user/customer feedback into a Feedback Analysis Pack (taxonomy, themes, recommendations). See also: conducting-user-interviews (collect new), designing-surveys (quantitative).

Apply behavioral science to product design: target behavior, intervention map, experiment plan.

Create a Brand Storytelling Pack: brand narrative, founder story, channel scripts, Q&A bank.

Build a promotion case: evidence log, impact narrative, manager alignment plan, submission-ready promo doc.

Build an early sales team: hiring plan, role scorecards, interview loop, onboarding/ramp.

Build or refresh a team culture code, norms, and reinforcement plan.

Judge whether any iterative loop is well-designed — a verification loop, review loop, maker/checker loop, evaluator-optimizer, or agent self-improvement loop — or design a new one from scratch. Provides a 10-dimension quality rubric (Mode A) and a 9-step build guide (Mode B) grounded in TDD, Reflexion, LLM-as-judge, process-supervision, and agentic-workflow research. Use when asked "is this loop good enough", "review/critique this loop", "design a loop to catch X", "what should a verification loop cover", or whenever you are designing, modifying, or evaluating the mechanics of a produce → check → fix → re-check cycle — even if the word "loop" is never said. Do NOT use for executing an existing concrete loop (run that loop itself), one-shot scripts, or plain Q&A.

Audit a repository's control-plane health and produce an actionable improvement report. Use when the user asks to "audit the repo", "check repo health", "assess control plane", "entropy audit for the control system", "check what's missing for agents", or wants to know what's missing in their instruction files, rules, and verification infrastructure. Do NOT use for code review (use review), dependency audit (use dependency-audit), general documentation work (use project-documentation), or reviewing a specific PR or diff (use review or entropy-review).

Review code changes for entropy introduction: naming drift, error-model splits, dependency violations, doc staleness, and pattern duplication. Activate when the user asks for "entropy review", "consistency check", "drift check on this PR", "check if this change introduces new patterns", or wants a review focused on naming, error handling, or structural consistency rather than correctness. Do NOT activate for general code review (use review), full repo audit (use repo-entropy-audit), control-plane audit (use control-plane-auditor), brainstorming, or implementation tasks.

Full-repository entropy scan across six axes: structure (dependency graph, layering, SCC), semantics (naming diversity, state model fragmentation), behavior (error model variants, boundary handling), context (doc freshness, instruction-file coverage), protocol (instruction-file consistency), and control (rule enforcement gaps). Produces a module-level heatmap, trend comparison, and prioritized cleanup targets. Activate when the user asks for "entropy audit", "repo health scan", "heatmap", "tech debt scan", or "full entropy report". Do NOT use for PR-level review (use entropy-review) or control-plane audit (use control-plane-auditor).

REQUIRED for Claude Code or Codex CLI hook queries — has accurate cross-platform reference. Triggers on PreToolUse, PostToolUse, Stop, SessionStart hooks, .claude/settings.json or .codex/hooks.json config, prompt vs command hook types, hooks not firing, exit codes, stdin JSON, matchers, path blocking, or dual-platform hook setups. Skip for React hooks, webhooks, GitHub Actions.

Analyze a coding project and design a tailored AI development environment: recommend agents, skills, instruction files, MCP servers, and workflow composition across Claude Code and Codex. Also audit an existing AI setup for gaps, overlap, or bloat. Use when setting up, overhauling, or auditing a project's AI tooling as a whole, not for individual agent edits or skill management.

Manage the lifecycle of sub-agents: create or update agent packages, validate cross-surface agent contracts, evaluate routing behavior, tighten invocation boundaries, tune Codex runtime defaults, install or publish agents, or audit an agent library. Use only when the request is explicitly about agents or agent libraries, not for ordinary coding, implementation, or general project planning.

Issue Agent OS controller loop. Pulls issues from a GitHub issue queue (priority + dependency resolved), fans out triage to parallel triager agents, dispatches coder / reviewer / splitter / planner / debugger workers based on verdicts, drives review loops, merges approved PRs, and repeats until the queue is empty. Use when the user asks to "process issues", "run the issue controller", "work through the queue", or similar. Do NOT use for single-issue ad-hoc work — this is the full automated loop.

Example Codex skill from Agentpack minimal repo template.

Operate agentpack safely in Claude Code. Use /ap-* commands for execution; prefer --json.

Contribute to Agentpack itself (spec-driven changes, tests, releases).

Operate agentpack safely (doctor/update/preview/diff/deploy/status/explain/evolve/rollback). Prefer --json.

聚合 agent_kaizen 两处收件箱的所有提案,裁决哪些是真问题、需要且能优化,并落位到合适的家。用于 review proposals、聚合审查、/review-proposals 时。

在 agent 干活时把「会重复的坑」或「可固化的优化」低打扰地记成一条 agent_kaizen 优化提案。用于发现优化点、想记一条提案、/propose 时。