Natural Writing

Natural writing

Rewrite prose in a natural, direct voice. Preserve meaning, facts, citations, numbers and intent while removing filler, canned framing and chatbot residue.

Hub role

This skill is the writing family's hub and owns academic prose directly. It holds the shared prose doctrine, anti-AI taxonomy, condense pass, claim-discipline schema, voice matching and the Australian English default. engineering-writing and legal-writing link here rather than clone it. Load au-english.md as the default house style unless a project convention overrides it.

Boundary

Use this skill for general, mixed or unclassified prose including correspondence, and for academic prose: thesis chapters, papers, abstracts and literature reviews. Defer technical docs and READMEs to engineering-writing, and Australian legal drafting to legal-writing. Correspondence on a specialist matter stays here, with that specialist as companion.

Do not fabricate experience, invent sources, add fake messiness or optimise for AI-detector scores. Formal prose is not proof of AI authorship.

Modes

• rewrite (default): clean the full draft.
• light-touch: preserve wording; repair only clear defects.
• match-voice: follow a supplied writing sample.
• precision-preserving: protect high-stakes or citation-heavy wording.
• full-rewrite: reshape low-stakes prose when a normal pass fails.
• diagnose: identify the strongest problems before rewriting.
• citation-safe, section-polish, final-scrub: academic variants.

Choose the least invasive mode; combine any with match-voice given a sample.

Workflow

1. Lock facts, logic, stance, citations, numbers, names, quotes and terms.
2. Diagnose the real defect: weak information, assistant residue, inflated language, repeated templates, unclear sentences or voice mismatch.
3. Rewrite structure, emphasis and rhythm rather than swapping synonyms. Use specifics already in the source; never invent texture.
4. Match confidence to evidence. Narrow or flag an unsupported claim, never strengthen it.
5. Repair remaining quality and voice mismatches, then check the result against the locked invariants and any sample.

Load patterns.md for the sign catalogue, voice-matching rubric or genre guidance, anti-ai-taxonomy.md for the tiered sweep, condense-pass.md for shortening, and claim-discipline.md for observed, inferred, limitation and pending claims. For academic prose load academic-prose.md: register, citation-key preservation, evidence altitude and thesis AI tells. LaTeX command and layout invariants live in engineering-writing.

Bright-line rules

• Preserve facts, meaning, citations, citation keys, named entities and quoted terms; flag a missing source rather than inventing a key.
• Delete chatbot framing and tool residue instead of paraphrasing them.
• Prefer concrete nouns and verbs; state each supported point once.
• Repair repeated templates without manufacturing quirks.
• Keep one honest qualifier for weak evidence; do not hedge strong evidence.
• If the draft already reads naturally, edit lightly and say so.

Output

Return the final text first, with notes only for material flags or choices. For a diagnosis, lead with the strongest risks. Keep inline flags short, such as [FLAG: cite source].