mimesis-human

mimesis-human

Humanise writing and strip the modern AI tells. The guidelines outrank the model: quality is a property of the reference corpus and the linter, not of whichever model is writing.

The precedence stack

Every run goes through three layers, top to bottom. The lower a layer sits, the more optional it is.

• L1, the guideline corpus. Read reference/tells.md and reference/craft.md on every run before doing anything. Primary, model-independent, never skipped.
• L2, the deterministic linter. Run the linter on the input, and again on anything you produce. It catches em dashes, smart quotes, kill-list words, negative parallelism and participle tails regardless of which model wrote the text.
• L3, the generator (optional). Claude by default. Codex / GPT-5.5 only if installed and opted into via --gen codex. See codex.md. Whatever generates, L1 and L2 always run afterwards.

Run the linter like this, against a file or stdin:

python3 "$CLAUDE_PLUGIN_ROOT/linter" --json path/to/draft.md

Use --json when you need to act on findings, or omit it for a readable report. The linter reports; it never rewrites. Restructuring is your job under L1.

Three modes

Detect the mode from the user's phrasing. Default to audit when ambiguous, so the skill never edits text the user only wanted checked.

• audit. Read-only. Run L1 and L2, report every tell with location, category and severity. No rewriting. This is the diagnostic pass.
• rewrite. Meaning-preserving humanisation of existing text. Apply the craft principles and the final pass from craft.md. Never change a factual claim, never invent experience or anecdotes.
• generate. Draft new text that is already humanised, rather than writing clean prose and cleaning it twice.

--mode overrides detection.

Two absolute rules

1. Zero em dashes, ever. When the linter flags one, restructure into commas, full stops, parentheses or two sentences. A restructure, not a swap. This also covers the en-dash-as-em-dash, double-hyphen and spaced-hyphen workarounds.
2. Catch the merely perceived. Flag and fix constructions that read as AI even when grammatically fine: tricolons, participle tails, dutiful sign-offs, antithesis. These get rewritten or downgraded, not waved through on a technicality. The linter handles the regexable ones; you handle the rest using tells.md and the final-pass checklist.

How a run works

1. Read tells.md and craft.md.
2. Run the linter on the input.
3. Mode audit: stop here and emit the report.
4. Mode rewrite or generate: produce the text, applying the craft principles and the final pass. If --gen codex and Codex is present, route the draft through it per codex.md, then continue.
5. Run the linter on your output. If it returns any finding, fix and re-run until it is clean. Never hand back text the linter still flags.

Output

audit. A tight report. No preamble, no sign-off.

# Humanise audit: <short description>

**Tells:** <n> blocker, <n> important, <n> polish

[1] line 4 (blocker) em dash: restructure into two sentences
[2] line 4 (important) negative parallelism: "it's not X, it's Y"
[3] line 9 (important) kill-list "leverage": use a plain verb

One line per finding, highest severity first. Lead with the em dashes; they are non-negotiable. Group the L2 findings, then add any [judgement] tells from tells.md the linter cannot see (flat rhythm, dutiful closing, neutrality).

rewrite / generate. Return the cleaned text and nothing else, unless the user asked what changed. If they did, follow with a short bullet list of the categories you addressed, not a line-by-line diff. End the turn on a clean linter pass.

Limits to state when relevant

• Detectors false-positive. This skill writes for signal and voice. It does not promise to beat any given detector.
• Meaning preservation is mandatory in rewrite mode. No claim changes, no fabricated anecdotes.
• The kill list is versioned, not eternal. It drifts as models drift.

What this skill never does

• Edit text in audit mode.
• Change a factual claim or invent experience to sound human.
• Swap an em dash for a comma without restructuring the sentence.
• Hand back output the linter still flags.
• Auto-route to Codex. Codex runs only on an explicit --gen codex.