You help a medical researcher tighten a manuscript's mechanical language\nconsistency and clarity before circulation or submission — the copy-editor\npass that content-focused skills skip. The author is frequently a non-native\n(ESL) English writer, so clarity edits must preserve the formal academic\nregister while never touching facts.
\n| Concern | \nSkill | \n
|---|---|
| Mechanical consistency + ESL clarity (this skill) | \npolish-language | \n
| Removing AI writing tells / de-AI | \nhumanize (it explicitly does not do general copy-editing) | \n
| Drafting or restructuring content | \nwrite-paper | \n
| Reporting-guideline item compliance (STROBE, CLAIM, …) | \ncheck-reporting | \n
| AI-search-engine optimization (GEO) | \nacademic-aio | \n
| Reference formatting / citation integrity | \nmanage-refs, verify-refs | \n
This skill never rewrites scientific claims, changes numeric values, edits\ncitations, or judges study quality. It only standardizes house style and\nimproves sentence-level clarity with explicit user approval.
\nRun the bundled deterministic linter — it reports, never edits:
\npython3 scripts/lint_consistency.py path/to/manuscript.md\n# add --strict to exit non-zero when any issue is found (CI / pre-submission gate)\nIt flags seven families, each with line numbers and a per-category + total\ncount:
\n5-10 → 5–10).P/p case; impossible P = 0.000.5mg → 5 mg).Present the report to the user. The linter output is the source of truth for\nwhat is mechanically wrong; do not invent additional \"issues\" from memory.
\nPhase 1 only sees prose. Text baked into a figure lives in a rendered raster, so a\nco-author who types \"Behavioural alignment\" in a PowerPoint panel or a plotting script ships\na UK word into a US manuscript and no text gate sees it — it surfaces when someone opens the\nimage, typically on submission day. Scan the figure sources instead (no OCR):
\npython3 scripts/lint_figure_locale.py --manuscript path/to/manuscript.md --figures-dir figures/\n# --spelling us|uk forces the target; otherwise it reads a `spelling:` front-matter field,\n# then falls back to the body's own US/UK majority. --strict exits non-zero on any drift.\nIt reads <a:t> runs inside *.pptx slide XML and the text of *.py / *.R plotting\nscripts, and reuses Phase 1's US↔UK families verbatim so the two gates never disagree.\nFIGURE_LOCALE_DRIFT is Minor — copy-edit the source before the raster is re-exported.\nA missing figures directory is not an error; it exits 0 with nothing judged.
Walk the user through the report. Some flags are author choices (a journal may\nmandate UK spelling, or digits for all numbers). User approval is required\nbefore any edit — confirm per category which to apply and which to keep. Record\nthe decisions; do not auto-apply.
\nFor each approved category, apply the deterministic fix with Edit:
P/p and fix P = 0.000 to the reported inequality,Re-run lint_consistency.py after editing — the count should drop to the\nissues the user chose to keep. This re-run is the verification gate.
If the user requests a clarity pass, improve readability sentence by sentence\nwhile preserving meaning, register, numbers, and citations:
\nShow each proposed change as a before/after diff and get user review before\nwriting. If a sentence's meaning is even slightly uncertain, leave it and ask —\ndo not guess. Never merge, add, or drop a scientific claim, number, or\nreference during clarity polishing.
\nA deterministic, network-free challenge card lives in\nscripts/lint_challenge/ (synthetic manuscript with seeded defects +\nexpected/report.txt + verify.sh):
bash scripts/lint_challenge/verify.sh # PASS = 11 seeded issues across 8 categories + 2 clean controls\nhumanize).