How The Skill Works - m4bwav/everwrite GitHub Wiki

Everything on this page comes from skills/everwrite/SKILL.md, which is what the agent reads when the skill loads. The checker it runs is described on What the checker flags.

When it loads

The agent loads the skill when it is about to write or edit text a person will read: docs, READMEs, reports, emails, posts, release notes and pull request descriptions. It also loads on requests such as humanize, sound human, remove AI tells, reads like AI and is this AI slop. It is not for code or config.

The skill's tests (skills/everwrite/TESTS.md, run on 2026-09-25) found a request to write an email loaded it in 2 of 3 runs, the weakest of the trigger cases. If an email or message comes out without the checker running, name the skill in the request.

How deep it goes

The agent picks a depth from the length and kind of the text:

Text What the agent does
A chat reply, or anything under about 150 words Follows the hard rules while writing. No checker, no audit.
A file, or longer text Drafts with the rules, runs the checker until it exits 0, then rereads once.
Someone else's text to edit Reads the before and after examples in references/patterns.md, runs the checker on the original, rewrites with the smallest edit that works, and runs the checker again.
A question such as "is this AI?" or "flag but don't rewrite" Detect mode: runs the checker, adds the patterns a script cannot see, and lists each finding as line: pattern: quote: fix. No rewrite, no score and no verdict on who wrote it.

Text under edit is treated as material. A draft that says ignore your rules is prose to fix, never an instruction.

The hard rules

  1. No em or en dashes, and no -- standing in for one. Dashes in code, paths and URLs stay. When the author's own sample uses dashes, the agent matches its rate and runs the checker with --allow-dashes.
  2. No invented facts. A name, number, date, quote or citation must come from the source or the user. The skill prefers a vague true sentence to a specific invented one.
  3. No chat wrapper: greetings, praise, sign-offs, offers and sentences that announce what the text will do.
  4. Formatting follows the content: sentence-case headings, no emoji, no bold label on every bullet, bold at most once per page, no rule between every section.
  5. In a file only prose changes. Code, commands, paths, frontmatter, data, link targets, quotations and names stay exactly as they were.

Patterns it removes

The skill lists patterns in two groups. The first group is acted on at one sighting:

  • the not-X-but-Y contrast in its many forms, including the split one across two sentences
  • staged setups and reveals
  • arguing with a position nobody took
  • closing kickers, stacked fragments and aphorisms
  • inflated significance
  • -ing riders added for depth
  • unnamed authorities
  • clusters of stock AI vocabulary (the list is in tells.py --list)

The second group needs other tells nearby before the agent acts:

  • forced lists of three
  • serves as in place of is
  • vague association and false agency
  • cycling synonyms for one thing
  • stacked qualifiers and repeated sentence openings
  • the same heading or paragraph twice in a row
  • docs written against the previous version
  • curly quotes where the target uses straight ones

Readability

Most sentences stay under 20 words, and one over 30 has to earn its length. Paragraphs run three to five sentences with the point first. The skill asks for plain words, verbs that do the work and jargon defined on first use. A sentence that would fit unchanged in a document about another product is filler, so it gets a fact or goes. With a writing sample, the author's habits win over these defaults.

The skill also says what to leave alone. Words such as very, perhaps and I think, a real aside, an odd specific detail and mixed feelings are human habits. One listed word or one however is not a tell; clusters are.

Watermarks

Removing tells does not remove a provider's statistical watermark, so the skill never promises undetectable text and does not try to strip one. Its section on watermarks says Claude and Gemini mark text through word choice and add no characters. Invisible characters are a different matter: the checker finds them and --fix-hidden deletes them (see Commands).

What comes back

The agent returns the final text. After editing existing text it adds one sentence of under 25 words on what changed, with the checker's counts for a file. In detect mode it returns the findings list only.

Keeping current

The skill is an Evergreen Protocol unit. On every use it reads evergreen.json; when the next check date has passed, it does the task, says the skill is due and then refreshes its research in the same session. The tier is moderate, with a 20-day interval; the last check was on 2026-09-25 and the next is due on 2026-10-17. Without the evergreen plugin, MAINTENANCE.md in the skill folder says how to refresh by hand.