Polish prose

Polish Prose

Turn an AI-flavored or overwrought draft into prose that reads like a careful researcher wrote it — without moving a single number, claim, or citation. Reviewers now pattern-match the LLM register (delve/leverage vocabulary, Moreover-stacked paragraphs, em-dash chains, uniform hedging), and a leftover "Certainly! Here is..." or "[insert citation]" is a desk-reject-grade embarrassment. This skill finds all of it deterministically, then guides a disciplined human-in-the-loop rewrite: expression changes, content does not.

When to use

• "My paper sounds like ChatGPT wrote it" / "humanize this" / "de-AI-ify"
• "Remove the AI words" / "it keeps saying delve and leverage"
• "Tighten / polish my writing" / "this is too wordy"
• "Fix my hedging" — overclaimed contributions or drowned-in-maybes results
• "Make my contribution statements active" ("a method is proposed..." → "We propose...")
• "Is my terminology consistent?" (dataset vs data set, acronym discipline, British vs American spelling)
• Called before preflight-check as part of the final submission pass, or after heavy drafting with any LLM assistant.

Inputs

1. The draft: a .tex file (preamble/math/verbatim are masked automatically), a .md/.txt file, or pasted text via stdin.
2. Optional: the venue profile venues/conferences/<venue>-<year>.yml (schema in venues/schema.yml) — supplies the venue family for register norms and review.llm_policy for AI-use disclosure duties.
3. Optional: a project glossary (canonical term = variant | variant per line) if the team has already standardized terminology.

Process

1. Freeze the technical content first. Build the no-touch inventory before editing anything: every number and unit, every \cite key, every stated result, dataset name, and system name. Snapshot it:

   grep -oE '[0-9][0-9.,]*\s*(%|\\%|ms|s|GB|MB|x|×)?' draft.tex | sort | uniq -c > /tmp/numbers-before.txt
   

   The same command must produce identical output after the edit pass (step 9). If polishing would require changing a claim, stop and tell the user — that is a content decision, not a style edit.

2. Resolve the venue register and the LLM policy. If a venue profile is given, read its family (register norms per references/venue-register.md) and review.llm_policy. Re-verify the LLM policy against the live cfp_url before relying on it — these policies churn every cycle. Tell the user plainly: polishing AI-assisted text does not remove a disclosure obligation. If the venue requires an AI-use statement, help write an honest one; never frame this skill as a way to avoid it.

3. Run the linters (Python 3, stdlib only, no network):

   python3 scripts/prose_lint.py main.tex
   python3 scripts/terminology_check.py main.tex --allow GPS,SDK
   

   Both accept .tex, plain text, or - for stdin; --json for machine-readable output; --strict to exit 1 while findings remain. Exit codes: 0 ok, 1 strict-failure, 2 bad input. --allow skips the undefined-acronym check for acronyms your community treats as universally known; --glossary FILE enforces a terminology table. Findings are candidates with line numbers, never auto-replacements — domain terms can look like tells ("leverage scores", "robust statistics").

4. Kill RISK findings immediately. Chatbot artifacts ("As an AI model...", "Certainly! Here is...", "[insert X]") are deleted or filled in, full stop. Several venues treat undisclosed LLM output as a desk-reject trigger; an unfilled placeholder is also a missing citation — route real citation needs through verify-citations, never invent one.

5. Edit pass A — lexical and structural tells. Work section by section with references/llm-tells.md open. For each WARN: keep it if it is a domain term or genuinely the best word (occasional use is fine — density is the tell, not existence); otherwise apply the rewrite pattern. Break Moreover/Furthermore paragraph chains by deleting the connective — if the logic no longer flows, the paragraph order was wrong, which is worth knowing. Recast surplus em-dashes as separate sentences or commas. Propose edits as diffs (old → new, with line numbers) and apply only with the user's approval.

6. Edit pass B — hedging and claims. Calibrate per references/hedging-and-claims.md: contributions, definitions, and completed actions lose their hedges ("we propose", not "we attempt to propose"); interpretations, generalizations, and limitations keep exactly one honest hedge; double hedges collapse to one; boosters (clearly, undoubtedly) are cut — the evidence carries the claim or nothing does. Rewrite passive contribution statements as active, specific claims that say what is new. Never strengthen a claim beyond what the paper's own evidence supports, and never weaken a result the user measured.

7. Edit pass C — terminology. Turn every compound-variant, spelling-mix, and acronym finding into a one-line decision: present the variants with counts, let the user pick the canonical form (default: the majority form; spelling per the venue register), then apply it everywhere. Record decisions in a glossary file and re-run:

   python3 scripts/terminology_check.py main.tex --glossary glossary.txt
   

   Fix acronym discipline: define once at first use, expand consistently, drop definitions never used again.

8. Edit pass D — venue register. Apply the family norms from references/venue-register.md: person and voice conventions, tense discipline, spelling convention, contraction policy, and the register differences between ML, systems, HCI, and LNCS venues. When no profile is given, default to consistent American spelling and the conventions already dominant in the draft.

9. Verify nothing technical moved. Re-run both linters (clean, or each remaining finding consciously accepted by the user); re-run the step-1 number snapshot and diff it — any difference is a bug in the edit pass and must be reverted; confirm the \cite count and keys are unchanged. Present a short before/after table for every sentence whose claim strength changed in pass B so the user signs off on each one.

10. Optional gate for CI or pre-submission: prose_lint.py main.tex --strict exits 1 while any RISK/WARN remains — useful as a final check alongside preflight-check.

Output

• Two lint reports (Markdown or --json) with line-numbered findings.
• Proposed edits as diffs, applied only with approval, one pass at a time.
• A terminology decision table plus a reusable glossary.txt.
• A verification note: numbers/citations diff clean, claim-strength changes enumerated, remaining accepted findings listed.

What this skill is not

• Not an AI-detector evasion tool. The goal is prose quality and an honest register, not beating a classifier. If the user's actual goal is to hide LLM use from a venue that requires disclosure, decline that goal, show the venue's llm_policy, and offer to draft the disclosure instead.
• Not a content editor. It does not restructure arguments, add citations, or change what the paper claims — tailor-to-venue handles repositioning, write-abstract handles the abstract, verify-citations handles references.

Relationship to other skills

• write-abstract — abstract-specific structure and venue metadata norms.
• preflight-check — the desk-reject gate; run it after polishing.
• anonymize-paper — identity leaks; this skill does not touch them.
• verify-citations — any citation need surfaced here routes there.

Adapt to your discipline

The tell lexicons are field-agnostic; the register is not. Fork and adjust references/venue-register.md for your field's journals (many mandate British spelling or third-person voice), extend the --allow acronym list for your community's alphabet soup, and add discipline-specific tells to prose_lint.py's TELL_PATTERNS table.

Guardrails

• Never alter numbers, results, claims, citations, or technical statements; expression only. Verify with the step-9 diff, every time.
• Never auto-apply lexicon findings — every WARN is a judgment call and domain terms ("leverage scores") are innocent.
• Never fabricate text for "[insert X]" placeholders; ask the user for the real content or route citations through verify-citations.
• Never present the result as "undetectable" or "human-written"; never help evade an AI-use disclosure requirement — surface review.llm_policy and re-verify it against the live cfp_url.
• Never submit anything to any system on the user's behalf.
• Quote at most the flagged line in reports; never reproduce large portions of the paper in outputs.

Memory

This skill uses the shared .paper-memory/ convention in the user's paper directory (full spec: paper-memory-convention.md).

• At start: read .paper-memory/lessons.md and profile.yml. Use writing_preferences (style signature, spelling, voice) to set the target register, and lead with any recurring prose habits already recorded (e.g. "you habitually write passive contribution statements") so you watch for them first instead of re-explaining them.
• At end: append durable findings in the shared format - [YYYY-MM-DD] (polish-prose | <scope>) pattern -> recommendation (via reflect-and-improve's reflect_log.py append, which dedupes and dates). A tell or habit seen across drafts is recurring; a one-off fix is this-paper. Do not log routine one-shot edits.
• Create .paper-memory/ on demand if absent and offer to add it to the project .gitignore. It is local-only; never upload it or copy it into this repo.