菜单
Use BEFORE implementing anything non-trivial — when a request is broad, multi-step, ambiguous, or risky and needs a decision-complete plan. Runs explore-first research, an interview on genuine unknowns, writes and self-reviews one plan, then stops at an explicit approval gate.
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。
Activate when the user types `csw` (alone, or as a prefix like `csw <task>`), or asks to "run the loop" / work with full rigor. Runs the evidence-bound autonomous loop — bind a goal, plan if needed, then drive each criterion through test-first implementation, real manual QA, independent review, and paired cleanup until the completion oracle passes.
Use to run a rigorous multi-lane review before declaring work done — for any broad, risky, shared, security-sensitive, or release-facing change. Uses native scheduling for independent review lanes and gates on ALL of them passing with a single binding verdict.
Use to EXECUTE an approved plan with discipline. Resumes from durable .csw/ goal state, drives each task through test-first implementation, real manual QA, independent verification, and paired cleanup, and only finishes when the completion oracle passes. Bootstraps csw-plan first if no plan exists.
Use when work contains two or more INDEPENDENT subtasks that can run at once (parallel investigation, multi-file changes, fan-out research, multi-lane review). Uses native Copilot task/fleet scheduling and integrates worker results skeptically.
基于 SOC 职业分类
| name | csw-plan |
| description | Use BEFORE implementing anything non-trivial — when a request is broad, multi-step, ambiguous, or risky and needs a decision-complete plan. Runs explore-first research, an interview on genuine unknowns, writes and self-reviews one plan, then stops at an explicit approval gate. |
Produce ONE decision-complete plan. Do not write product code in this skill.
State the request's shape: one-file change · multi-step feature · architectural decision. Trivial one-liners do not need this skill; say so and stop.
Ground every claim in the actual codebase and sources. Parallelize independent investigation with the swarm:
task → explorer — where things live, how they connect,
path:line evidence.task → researcher — library/API facts, each pinned to a commit
SHA / permalink.
Use /fleet for user-visible parallel execution and /tasks for oversight or
cancellation. Before investigation, use host deny/available-tool policy to remove
mutating tools; prose alone does not enforce non-mutation.
Re-read the cited evidence yourself; do not trust a worker's summary.Ask the user ONLY about decisions you cannot resolve from the request, the code, or sensible defaults — and where the answer changes the plan. Do not ask about things you can verify. Present materially different interpretations rather than silently picking one.
Write a single plan file under plans/. It must be decision-complete:
C0NN | channel: | test: | scenario:) the goal runtime
can ingest (bin/csw-runtime.mjs init).Delegate two analysis-only passes and fold their findings back in. Enforce non-mutation with host tool policy before launch:
task → gap-analyst — contradictions, ambiguities, missing constraints, risks.task → plan-reviewer — task granularity, file existence, testable criteria,
concrete QA scenarios. Verdict must be APPROVE (loop on ITERATE/REJECT).Present the reviewed plan, findings, recommended approach, remaining decisions,
and scope (must-have / must-NOT). Wait for the user's explicit approval. Never
auto-proceed from plan to implementation. Only after approval may the plan hand
off to csw-work.