Skill v1.0.0
currentLLM-judged scan90/100version: "1.0.0" name: writing description: 'Writing prose for human readers. Load it before drafting, editing or reviewing any of these, however short: an email, chat message or reply the owner sends; a README or docs page; a technical explanation, plan, report or spec; or a blog post, landing page or announcement. Not for your own chat replies to the user.'
Writing
One skill for prose that people read. Pick the mode by who reads the text and what they do after reading it, read that mode's file, and apply the shared rules below to every mode.
Before returning any text, read the AI-tell catalogue in `../humanizer/SKILL.md` (~/.agents/skills/humanizer/SKILL.md) and check both the text and your chat reply around it against its sections. For docs, technical and marketing text, read the catalogue by itself and in full, with the Read tool or by loading the humanizer skill: never in one command with other files, never through head or a grep of its headings.
These five tells most often survive a rewrite:
- §1 Not X but Y: a contrast whose negative half denies something nobody claimed.
- §2 Closers: a last line or fragment that repeats the point instead of adding to it.
- §6 Triads: three items or examples where the meaning has fewer parts.
- §8 Dashes: an em or en dash outside code, where a period, comma, colon or parentheses fits.
- §19 Bold labels: list items that each open with a bold label and a colon, or bold that marks nothing.
Route
| Reader, and what the text is for | Read | |
|---|---|---|
| Someone you write to as the owner: an email, chat message, reply or comment | communication.md | |
| People using or building on a product: a README, docs page, guide, reference or changelog | docs.md | |
| A colleague or the owner who must understand or decide: an explanation, plan, report, spec or postmortem | technical.md | |
| A public audience you want to win over: a blog post, landing page, announcement or launch post | marketing.md | |
An agent: an AGENTS.md, CLAUDE.md, skill, subagent brief or prompt | the writing-for-agents skill instead of this one |
When two rows fit (an announcement sent as an email), pick by the reader's next action. Pasted text that only needs its AI tells removed goes straight to the humanizer skill.
Done when you have read exactly one mode file, or handed off.
Shared rules
- The point comes first. Name the reader and what they must know or do after reading. The answer, ask or decision opens the text; supporting detail follows.
- Every sentence adds something the reader lacks. Cut background they already have, restatements, and announcements of what comes next.
- Facts trace to a source. Every name, number, date, quote, commitment and claim comes from the user, the repo, or something you read or ran. A gap becomes a question or a visible placeholder (
[confirm: launch date]), never a plausible guess. - Specific beats general. Numbers with units over adjectives, the real symbol, file or command over a description of it, one concrete example over an abstraction.
- Match the voice that already exists. A writing sample, the thread you are replying in, or the repo's existing docs set sentence length, punctuation and terms. Use the repo's glossary terms when it has a glossary. With none of these, write plain and direct.
- Finish with the AI-tell pass. Check your draft against the catalogue linked above: mark the tells strongest first, rewrite the sentences around them, then search the text and your reply again for the five tells above, and confirm the rewrite added or dropped no fact. Done when a full read of the text and the reply finds no §1 to §5 tell, no §19 bold label and no dash outside code.