polish

name	polish
description	Polish prose in three passes: strip AI-writing tells, apply Strunk's rules for clarity, then hold every research claim to its evidence. Two modes: edit in place, or diff-first where the original is untouched until the user approves each change. Use when the user asks to 润色 / 改写 / 精修 / 打磨 a draft, or says "polish", "refine", "tighten", or "improve". Preserves meaning, numbers, citations, and [NEEDS-EVIDENCE] markers; never smooths an unsupported claim into a supported-sounding one.
license	MIT
compatibility	claude-code
allowed-tools	["Read","Write","Edit","Bash","Glob","Grep","AskUserQuestion"]

Polish

Polish runs three passes over a piece of prose, in this order:

1. De-AI, strip the marks of machine-written text (inflated symbolism, em-dash overuse, rule-of-three, promotional language, vague attribution, causal-tail chains, filler). Read references/ai-tells.md.
2. Strunk, cut needless words, prefer active voice, prefer the concrete over the abstract, keep related words together. Read references/strunk.md.
3. Evidence-led wording, for research and systems-paper prose, hold every comparison, magnitude word, and novelty claim to its evidence, and preserve unresolved-evidence markers. Read references/evidence-led-wording.md.

The order matters. De-AI works on texture: cadence, register, cliché, AI-brand punctuation. It does not shorten sentences; it makes them sound less like a chatbot. Run Strunk first and it will happily shorten AI-inflected prose into shorter AI-inflected prose, clipped but still hollow. Strunk second lands its cuts on real content instead of on decoration. Evidence-led wording runs last so that neither earlier pass has quietly turned an unknown fact into an asserted one while smoothing the prose.

The output is a single revised version, not an explanation, not a checklist of tells you caught, unless the user asked for one or chose diff-first mode.

Two modes: pick one before you start

In place is the default for a pasted paragraph, a single section, or any time the user just wants the better version back. Apply the three passes and return the revised text. If the input was a file, Edit it directly.

Diff-first (--diff, or when the user says "show me the diff first", "每一处改动先给我看", "先给 diff 再决定", "don't touch the files until I approve") never edits the original. You polish a copy, present each change as a unified diff, the user accepts, rejects, or partially accepts each one, and only then do you apply the approved subset. See the procedure below.

Pick diff-first by default, even unasked, when the document is close to submission (a paper, rebuttal, or camera-ready spanning multiple section files). The stakes are high enough that a silent over-edit is worse than the review overhead. Pick in place for everything smaller.

This is the one place a strong model should not improvise the orchestration: do not spawn parallel subagents to polish sections. Work the sections yourself, one at a time. The per-change review is the real control gate, not fan-out speed.

The preservation contract (both modes)

Polish preserves meaning at the claim level, not the sentence level. Every factual claim, every hedge the author meant to keep, every named entity, citation, and number must survive. Polish must not silently change the argument.

Byte-exact, never rewritten:

• All markup: \cite{...}, \ref{...}, \label{...}, captions, inline code, links, footnotes, % claim: id line comments. Rewrite only the surrounding prose.
• Every measured number, percentage, unit, and uncertainty statement. Do not round, restate, or recompute.
• Sentence-level claim boundaries. Do not collapse two % claim:-tagged sentences into one.
• [NEEDS-EVIDENCE] markers. This is the load-bearing invariant. The marker is an unresolved claim, not awkward prose. Preserve it verbatim, attached to the same claim. Never replace it with a vague hedge, an unrelated number, a section reference, or smoother, more confident wording. Never invent a comparator, measurement, citation, figure, or table to make the sentence read cleanly. If the evidence is absent, keep the marker or return a neutral description that does not make the unsupported claim.

When clarity conflicts with evidential precision, keep the precision. A smoother sentence that converts an unknown fact into an asserted fact is a regression, not a polish. The third pass (references/evidence-led-wording.md) is the detailed contract for faster, significant, efficient, scalable, and novel; run it on research prose and honor it.

When does the evidence pass fully engage? When .writing/metadata.yaml has writing_profile: systems, when the input is a LaTeX research manuscript, or when the user says the prose is a paper. For an email or a blog post it is a light touch, but the [NEEDS-EVIDENCE] rule above holds for any input that carries the marker.

Canonical-term table (consistency anchor)

Before polishing anything longer than a paragraph, build a canonical-term table. List each load-bearing term once with its single preferred surface form, then hold every rewrite to that form. This stops the de-AI pass from synonym-cycling a fixed term and stops the Strunk pass from collapsing a precise term into a bland one. See skills/_shared/core/terminology-ledger.md for the format. For a short input a mental table is enough; for a section or longer, write it down so it survives chunking.

Diff-first procedure

Step 1 : Find the sections, skip stubs

For a LaTeX manuscript, list the section files in deterministic order (default .writing/manuscript/*.tex). Skim each before starting. A stub is a tiny file whose body is a % redirect ... comment or an \input{...} shim; skip it, polishing it is wasted work. If the user named specific sections, do only those.

Proceed decisively when discovery is unambiguous. If the path is named or ls returns a clear set of section files, state the list and move on. The per-change review is the control gate; a redundant "shall I proceed?" just adds friction. Ask only when discovery is genuinely ambiguous.

Step 2 : Polish a copy, never the original

Work section by section. For each, copy the original, run the three passes on the copy, and generate a unified diff against the original. A convenient home for the copies and diffs is .writing/polish-patches/ (create it if missing), but for a single pasted section you may hold the diff in memory and present it inline. The invariant is only that the original file stays byte-unchanged until Step 4.

If a project style file (.claude/CLAUDE.md, AGENTS.md) sets prose rules, treat them as hard constraints that override default polish heuristics, and surface any rule the polish would have violated.

Step 3 : Walk the diffs with the user, one change at a time

Present a short per-section summary (line count, change count), then walk every change individually. Do not dump a long diff all at once.

Format for each change:

### Change N (paragraph location, brief subject)

```diff
- <full original prose, no ellipsis>
+ <full polished prose, no ellipsis>
```

判断: <one to three sentences: what changed and why>.
建议: 接受 / 拒收 / 部分接受 (具体哪些子改动), <reason>

Hard rules for this step:

• Always use triple-backtick diff fenced blocks. The terminal renders - red and + green; that visual signal is doing real work.
• Never use ... ellipsis in the diff payload. The user needs the full prose to judge context. Show it all, even when long.
• Walk every change, not just the surprising ones. The user is doing the audit; your job is to present, not to filter silently.
• Classify each as accept, reject, or partial. Partial-accept is common in a long change where the polish made four sub-edits and only two are clean; spell out which sub-tokens to take.

Step 4 : Apply the approved subset

After the user approves, apply via Edit with exact byte-strings. Do not bulk-apply via replace_all unless the change is genuinely the same string everywhere. If Edit rejects an old_string, the file diverged from the diff base (a linter or the user touched it mid-review). Re-Read, find the surviving form, and adjust. Never overwrite user edits silently.

Unapproved suggestions stay frozen on disk for later review. Report a short final tally: sections processed, changes applied, rejected, partially accepted.

Recurring over-edits: reject or accept by default

These patterns recur across manuscripts. In diff-first mode, flag them; in-place, just do not make the rejected ones.

Reject by default:

• Stripped parenthetical limiters at a term's first use. The de-AI pass reads parens as defensive, but they often set the limited quantifier of a defined term. Naive (per-hash) AOT becoming Naive per-hash AOT loses the scope.
• Modifier reordering that flips the argumentative subject. When a modifier moves to a different head, check the head is still the same noun. If not, reject.
• Deletion of editorial qualifiers that bind downstream. "and most consequential in production", "above all", "crucially", "the central case" are often load-bearing signals the next paragraph relies on. If it does, keep them.
• Semicolon-merge of two definition lines. When two short sentences introduce a defined term, the rhythm break is intentional. Keep them separate.
• Register downgrade. Swapping an academic copula for a breezy phrasing (is part of becoming falls out of). Keep the original.

Accept by default:

• Removal of self-deprecating qualifiers around real design choices (as an extension, albeit auxiliary, not a central experimental claim), especially when a later section states the trade-off confidently.
• Active-voice conversions that preserve meaning (is X by Y to Y X-es).
• Defensive-negation removal where the negation guards against a critique no reviewer would raise (rather than by an arbitrary modeling compromise).

When to use this skill

• "polish this", "clean this up", "make this tighter", "improve this paragraph".
• "润色", "打磨", "精修", "改写", "改得更好", "让这段更自然", "帮我改下这段".
• "整篇润色但每一处改动先给我看" (diff-first over a whole manuscript).
• A pasted draft with an implicit ask ("what do you think?", "thoughts?") when the prior turn made clear the user wants a revised version.
• Any request to edit prose for readability rather than factual correctness.

When NOT to use this skill

• Factual correction, translation, summary, or structural rewrite (reorganize sections, change the argument, add or remove content). Those are different jobs. Polish preserves meaning.
• LaTeX, code, tables, or bibliography fixes. Polish is for prose.
• A single sentence the user wants nudged. The three-pass machinery is overkill; edit inline and say you did.
• The user already rejected a de-AI or Strunk change earlier in the conversation. Respect that; do not re-apply it silently.
• "Don't change meaning, just fix grammar." That is a grammar pass, not a polish.

Edge cases

• Mixed Chinese / English prose. The passes apply cleanly to English. For Chinese, keep the same spirit (cut AI tells, cut needless words, prefer active voice, prefer concrete nouns) rather than mechanically applying English punctuation rules. Never translate or romanize either language.
• Very long input. Process one logical chunk at a time (paragraph or subsection). Running all three passes over thousands of words in one shot produces uneven quality, and chunking preserves headings, lists, and code blocks by touching only prose.
• Register mismatch after the de-AI pass. If it sounds too casual for the document (it stripped academic register from a Methods section), dial it back on the next attempt rather than letting Strunk compound the drift. Prefer one good polish over iterative grinding.
• Academic input is often near-identity. A clean systems-paper paragraph may come back unchanged, and that is the correct result. Hyphenated compounds (cold-cache, per-contract) and scholarly verbs (shows, preserving) are load-bearing; if a pass touched them, the preservation contract restores them.
• A polish that renames a \subsection{...} or section title. That is a semantic edit, not a stylistic one. Surface it separately with explicit confirmation; never bundle it into a prose batch.

Output philosophy

The polished text is the deliverable. The user does not need to see the machinery. In place, return the revised prose and nothing else, unless one short note earns its place (for example, "I kept the hedge in the third sentence because it carries the sample-size claim"). Meta-commentary about which tells you removed does not earn its place.

In diff-first mode, keep user-facing text focused on the change under review: which one, what changed, accept / reject / partial, a one-sentence reason. Do not narrate the pipeline or the patterns being filtered.

If a pass produced a result you are not confident in, say so briefly so the user can spot-check. Do not hedge for the sake of hedging, but be honest about low-confidence edits.