菜单
Post-implementation triage: measure actual diff metrics to determine review depth. Run after implement, before review. Usage: /estimate (in a worktree with changes)
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
Sprint lifecycle: plan, run, review, retro. The top-level orchestrator for autonomous issue resolution. Use for "run a sprint", "plan the sprint", "sprint review", "sprint retro", "/sprint", "/sprint plan", "/sprint review", "/sprint retro", or any variant.
Step back and interrogate the frame instead of executing it. `/why <subject>` GENERATES the sharp questions that should have been asked about a request, plan, artifact, or your own intended action — it does NOT answer them. Use when handed an assignment that smells mis-scoped, when about to loosen a constraint to silence a signal (bump a timeout, widen a threshold, disable a check, narrow scope to the measurable half), when triaging borderline/unfiltered backlog issues, or when a user is frustrated that the obvious framing is wrong. Triggers: /why, /why nsfw.
Build an autonomous sprint skill for a target project. Explores the project's workflows, constraints, and tooling, then produces a tailored sprint skill that lets Claude orchestrate parallel implementation sessions end-to-end. Use when setting up auto-sprint in a new repo, or when adapting sprint orchestration to a project with different workflows. Trigger on "bootstrap sprint", "set up auto-sprint", "build sprint skill for <project>", or "make <project> sprintable".
Audit and garbage-collect a committed Claude-memory store (MEMORY.md + per-fact files). Evidence-based, parallel, then a gated apply. Use when MEMORY.md has grown large, after several sprints, when memory has drifted from CLAUDE.md, or when the user says "/memory-gc", "clean up memory", "audit memory", "the memory is huge", or notices memory diffs/orphans.
Extract actionable lessons from sprint history. Buckets every Claude Code session into the sprint it ran in, builds a deterministic per-session digest (tokens, duration, errors, orchestration signals, evidence), then runs a multi-agent Workflow that classifies each session (orchestrator vs worker), scans for tool errors / autoclassifier rejections / user interventions, summarizes it, aggregates per-sprint, per-10-sprint window, and over time — ending in ranked, concrete actions. Use for "extract sprint lessons", "what can we learn from sprint history", "/sprint-lessons", or a retro deep-dive.
Author, refine, or migrate a doing-it-wrong architectural rule, and harvest recurring mistakes from merged PRs. Routes on its argument — `harvest` (mine PR review comments for recurring author mistakes), `init` (bootstrap the rule engine into a repo), `add` (author a rule), or no arg (overview + "is this a good rule?" triage). Use when the user wants to "write a rule", "add a lint rule", "should we lint for this", "make a rule for X", "migrate a check-*.ts into the rule engine", "harvest rules", "what should we lint for", "reduce review churn", "mine PR comments for rules", or judging whether a proposed invariant is worth a rule.
| name | estimate |
| description | Post-implementation triage: measure actual diff metrics to determine review depth. Run after implement, before review. Usage: /estimate (in a worktree with changes) |
Measure the actual implementation diff and decide how much review it needs. This replaces upfront estimation — the implementation itself is the best signal.
Predicting effort before implementation is unreliable (validated: best model achieved only 29.7% tier accuracy on 91 historical PRs). Instead:
After implementation completes, run triage against the PR (works from any directory):
bun .claude/skills/estimate/triage.ts --pr 1152 --json # PR-based (preferred)
bun .claude/skills/estimate/triage.ts # local diff vs main
bun .claude/skills/estimate/triage.ts --base develop # local diff vs develop
High scrutiny if ANY of (validated: 92.5% F1, 0% false negatives on 91 PRs):
Everything else is low scrutiny.
Low scrutiny: implement → QA → done
High scrutiny: implement → adversarial-review → QA → done
triage.ts — post-implementation triage (reads git diff, outputs scrutiny level)score.ts — AST-level complexity analysis (used by triage and backfill)backfill.ts — mines historical PRs into SQLite for validationembed.ts — text embeddings for PR similarity (research; not used in triage)validate.ts — method comparison (proves triage > prediction)validate-triage.ts — validates triage rules against historical datadb.ts — shared SQLite database helpers (used by backfill, validate, triage)Historical PR data lives in ~/.mcp-cli/estimates.db. Run bun .claude/skills/estimate/backfill.ts to populate/update.