菜单
Use when the user wants to analyze their Claude Code chat history for recurring patterns, forgotten threads, or skill-abstraction candidates ("review my chats", "abstract skills from chat logs", "what am I repeating", "find patterns in my Claude sessions", "retro my chat history").
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when the user wants Claude, Codex, agy, and the Cursor models Composer 2.5 and Grok Build to iterate together until they converge on a mutually-agreed result — or hit a deadlock the user must arbitrate — and for fresh-eyes adversarial review of drafted artifacts. Five modes — plan, implement, verify, review, and audit (the folded-in adversarial review). Triggers — "converge", "have claude and codex work it out", "iterate until you agree", "three-AI consensus", "multi-AI consensus", "deadlock me a decision", "adversarial review this", "audit my drafts", "fresh eyes on this", "review for fabrications".
Use when starting any substantial feature — the before/during/after development lifecycle: interview rounds, parallel blind-spot/references/reality-probe passes, a knowns matrix folded into the plan, an implementation-notes log during the build, and a compounding step after (explainer, quiz, memory/skill extraction). Triggers — "feature passes", "run the pre-build passes", "blind spot pass", "knowns matrix", "interview me about this feature", "compound the learnings".
Use before baking an external config value — a model id, endpoint, ARN, bucket, region, credential, API version, image tag — into code or a deploy. Invoke the real dependency ONCE against the target account/environment and read the response, so a dead/renamed/deprecated/permission-blocked value is caught at the source instead of failing in production. The config equivalent of "verify against reality, not the doc." Triggers — "probe before wire", "smoke the dependency first", "is this model/endpoint/ARN actually live", "verify the config before deploy", "does this id still work", "check it against the real account".
Use when an agent, workflow, or fan-out died mid-work — API connection drops, session/spend limits, forced structured-output timeouts — and work must be recovered without loss. Triggers — "resume the run", "the agent died", "hit the session limit", "recover the workflow", "connection closed mid-response".
Use when you need to maximally verify a piece of code — especially AI-generated code — by stacking every independent check it admits (differential/property/metamorphic/fuzz/concurrency/static/cross-model), anchored on the strongest available oracle, and breaking the correlated-failure trap (one agent wrote both code and tests). Triggers — "how do I maximally verify this", "make this bulletproof", "how should I test this", "make this safe to optimize/refactor with AI", "what oracle should I use", "I don't trust the agent's tests", "maximize verification", "is my test suite actually verifying anything".
Use when a deployed/wired integration is FAILING (silently or opaquely) and you're tempted to try a fix. The failure lives at one of several layers — your wiring, your env, your IAM, an org policy/SCP, an account-level enablement/subscription, or the provider's response format — and each layer has a DIFFERENT OWNER. Ladder down them, verify what you own is actually applied, and get to the ONE authoritative signal (the real log line, an elevated-creds probe, latency-as-discriminator) before attributing blame or attempting a fix. Triggers — "why is this silently failing", "it deployed but doesn't work", "ladder the failure", "localize before fixing", "read the real error", "is it my code or their config", "diagnose the integration", "AccessDenied but the IAM looks right".
| name | review-chats |
| description | Use when the user wants to analyze their Claude Code chat history for recurring patterns, forgotten threads, or skill-abstraction candidates ("review my chats", "abstract skills from chat logs", "what am I repeating", "find patterns in my Claude sessions", "retro my chat history"). |
Analyze Mike's Claude Code chat history (~133 project corpuses at ~/.claude/projects/) to surface recurring workflows that could become skills. Self-bootstrapping — uses parallel Explore subagents to avoid blowing up the main context.
PAIRED WITH: When a candidate is selected, chain into new-skill to scaffold it.
REQUIRED BACKGROUND: Uses superpowers:dispatching-parallel-agents pattern — up to 3 Explore agents in parallel, each with a disjoint slice of the corpus.
Use when the user asks:
Do NOT use when:
Each chat log is at ~/.claude/projects/<slug>/<session-uuid>.jsonl. Lines are JSON objects; user messages have "type":"user". The first user message of each session is the session-intent signal — most informative per byte.
Total scope (as of 2026-04): ~133 project dirs, mostly small. A handful are large (SignLab-Dev 331MB, superpowers 11MB, etc.) — weight sampling by corpus size.
Scope the corpus. Default: all of ~/.claude/projects/. If the user specifies a project filter ("in roxas") or date range ("last month"), narrow accordingly.
Partition the corpus into 3 clusters for parallel analysis:
du -sh)Dispatch 3 Explore subagents in parallel — one per cluster. Each agent's prompt:
Merge & dedupe the three reports. Patterns surfaced by 2+ agents are high-confidence.
Cross-check against installed skills. Run ls ~/.claude/skills/ && ls /Users/mikelady/dev/mike-skills/ and compare trigger phrases in candidate descriptions. Drop any candidate whose triggers heavily overlap with an existing skill.
Present candidates via AskUserQuestion (multi-select). Each option shows: pattern name, frequency, one example. Include an "add your own" by letting Mike reply with "Other".
For each selected candidate: chain into the new-skill skill to scaffold it.
Pass these hints to the subagents:
| Signal | What it means |
|---|---|
| Same opening phrase across ≥3 sessions (e.g., "install https://...") | Strong skill candidate |
| Repeated tool-use sequences (Bash → Grep → Bash) | Workflow worth abstracting |
| First user message is long and highly structured | Often a template the user re-types |
| Terms like "again", "like last time", "the usual" | User is doing repeat work |
| Sessions ending mid-stream (assistant tool calls, no user response) | Forgotten threads, not patterns |
Pass this into every subagent prompt so they can filter:
( ls ~/.claude/skills/ ; ls /Users/mikelady/dev/mike-skills/ ) | sort -u
After all subagents report:
Found N candidate patterns (M after filtering against installed skills):
1. <pattern-name> — <frequency>
Example: "<verbatim user message>"
Proposed scope: <one sentence>
2. ...
Then use AskUserQuestion to let Mike select which to build.
dialed:*, peon-ping-*, superpowers:* count — include them in the filter.If Mike has already run review-chats recently (check /Users/mikelady/dev/mike-skills/review-chats/last-run.md if present), start from that baseline rather than re-scanning the entire corpus. Ask Mike if he wants a delta from last run or a fresh full scan.