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.