humanize-and-unslop

Files (6)

• SKILL.md
• evals.json · evals/
• examples.md · references/
• judgment-patterns.md · references/
• patterns.json · references/
• scan.mjs · scripts/

SKILL.md (first lines)

# Humanize and unslop

Strip AI tells from prose while keeping the author's claims and voice. The model reads and rewrites. `scripts/scan.mjs` is a precise checker for words, phrases, and punctuation, used to confirm the result is clean.

## Rules that always apply

- Keep every claim. Never change facts, numbers, severity, scope, or meaning, since this is a style edit.
- Leave code blocks, inline code, URLs, commands, and quoted material untouched. The scanner already skips them.
- Use straight quotes (`"` and `'`), never curly quotes.
- Use no em dashes and no spaced en dashes. Use a period or comma instead, not parentheses.
- Use colons only to introduce a list or example, never as a lead-in ("The result: ...").
- Use sentence case for headings. Keep emojis and decorative bold out of headings, bullets, and body text.
- Return text that is already clean unchanged.

When rules pull in different directions, apply this order: preserve meaning, then fix scan findings, then make the smallest edit that works, then add soul.

## Workflow

1. **Resolve mode and input**
   - Use Detect mode when the request says `--detect`, "audit", "check", or "flag". Otherwise use Edit mode.
   - Count words. Under about 150 words, skip the scanner in steps 2 and 5 and check by reading.
   - For longer pasted text, write it to a temp file (from `mktemp`) so the scanner can read it. For a file path, scan the file directly.
   - In Edit mode, ask for audience and format only when the register is unclear and a person can answer. Otherwise infer the register from the draft.

2. **Scan**
   - Execute `node ${CLAUDE_SKILL_DIR}/scripts/scan.mjs <file>`.
   - Treat `fix` findings as required changes. Treat `review` findings as terms with a possible literal or technical sense (`primitive`, `surface`, `vector`); keep them when that sense applies.

3. **Judge**
   - Read `references/judgment-patterns.md` and check the draft for everything the scan does not cover: Title Case headings, colon lead-ins, bold-label bullets, binary contrasts, forced triads, synonym cycling, passive voice hiding the actor, uniform rhythm, unsourced claims.
   - Decide the register. Apply the "adding soul" section only to personal or opinion writing, never to docs, specs, READMEs, or third-person reports.

4. **Edit (Edit mode only)**
   - Read `references/examples.md` before the first edit in a session to calibrate edit depth.
   - Rewrite to clear every `fix` finding and each confirmed judgment finding.
   - For a file input, edit the file in place. Keep the change summary in the reply, never in the file.

5. **Verify**
   - Re-run `scan.mjs` on the edited text. Repeat step 4 until zero `fix` findings remain.
   - Reread once and fix whatever still reads as generated, usually rhythm or a summary ending.

## Output

- **Edit mode:** the edited draft (or a note that the file was edited), then a `## What changed` list of 3 to 6 bullets covering structural, tonal, and wording changes. Mention any `review` findings deliberately kept, and why.
- **Detect mode:** do not rewrite. List findings as `line: [pattern] "exact quote" -> brief fix`, scan findings first, then judgment findings, then a one-line count per pattern. No paragraphs of feedback.

## Reuse

Other skills and agents that need these rules should read `references/patterns.json` and `references/judgment-patterns.md` from this skill's installed folder (in the global or the project skills directory), or run `node <that folder>/scripts/scan.mjs --json <file>`, rather than copying the word lists.

From ersanyamarya/agentry-labs, MIT.

More from ersanyamarya/agentry-labs