cleanup-editor
Self-edit a draft for clarity, contradictions, answerable open questions. Use after initial draft.
- License 许可
- MIT
- 最近更新
- 2026-09-16
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- cleanup-editor
- description
- Self-edit a draft for clarity, contradictions, answerable open questions. Use after initial draft.
03 SKILL.md 正文
Just wrote a draft? Run yourself through this pass before handing off. Goal: clean linear story — final claims direct, contradictions resolved, open questions either answered or genuinely user-judgment.
Edit the draft in place.
Find + fix
Thought-process prose
Doc narrates your deliberation ("I considered X, then realized Y, but then thought about Z and decided…") instead of stating the final plan. Rewrite as "here is what we will do." Reader doesn't need the path.
"But wait" / "actually" reversals
Mid-prose stream-of-consciousness reversals where you discover your error and correct mid-paragraph. Drop the wrong half, keep the correct claim. Correct answer not obvious from surrounding text → contradiction (next item) — investigate, don't just delete.
Contradictions between sections
Section A says X; section B says ¬X. Find the truth (re-read code/spec/prior section). Rewrite both to agree. Don't paper over by hedging both — pick the right answer.
Open questions you can answer
Draft poses a question that investigation would resolve — read a file, grep, check schema, look up an API. Investigate, answer, update doc. Final draft's open questions = genuine user-judgment calls only (taste, prioritization, "what does the user want?"). Anything answerable through investigation is your job.
Approval requests masquerading as open questions
Cut on sight, every time: "confirm this direction", "approve X before proceeding", "the spec reserved this call so the user confirms here", "sign off that the earlier answer still holds", or anything else whose only answers are "yes, go ahead" and "no, change it". The doc already goes to a user gate — approval is the workflow, not a question, and asking for it burns the user's attention on a decision they were about to make anyway. If the reasoning eliminated every alternative, the author made the call: promote it to a stated decision with the reasoning that forced it, and delete the question. Also delete notes that merely observe an already-answered question stays answered. Pruning empties the section → remove the section.
Terminology drift
Same concept named two ways. Pick one, normalize. Can't tell if they mean the same thing → contradiction → investigate.
Promised but missing sections
Outline / earlier section promises something the doc doesn't deliver ("we'll cover X below…" with no X). Either write the content or remove the promise.
Speculative / unverified claims
"I think this should…", "probably this means…". Matters? Verify. Doesn't matter? Cut.
Filler
Throat-clearing intros, restated context the reader already has, "as mentioned above" cross-references earning nothing. Cut.
Preserve
- All substantive technical content. Don't trim detail to shorten — only trim process narration, redundancy, filler.
- Genuine user-judgment open questions.
- Any "Human-review delta" block explaining what changed since user last read.
- Domain terminology + your authorial voice on substance.
Operate
- Re-read draft top to bottom.
- Per issue: prose (rewrite in place) or substance (investigate + resolve)?
- Do prose edits.
- Do substance work — read what you need, update with the answer.
- Done: no thought-process prose, no contradictions, no answerable open questions, no promised-but-missing sections.
You're not creating a separate cleanup report. Improve your own draft directly.
04 同仓库更多 Skills
setup-project
Bootstrap a project for the review-chain workflow — set the orchestrator as default agent, add a "Working With Claude Code" section to CLAUDE.md, install the TODO.md + TODO(slug) tracking convention, and optionally wire an auto-format hook. Take it or leave it; the review-chain agents work without any of this.
usage-wall
Ride the 5-hour usage window into the wall — schedule a post-reset wakeup, then resume stalled subagents. No watchdog, no throttling.
usage-guard
Arm a background watchdog on the account's 5-hour usage window
simplify
Review changed code for reuse, quality, efficiency; fix issues found.
configure-models
Generate or edit local per-agent model overrides for review-chain agents (.claude/review-chain-models.conf) — re-pin a model without editing agent files or publishing.
clarify
Restate your most recent response more clearly and concisely