clear-and-human

# Clear and Human A writing skill in three layers: **construct** good prose (Strunk), **detect/score/report** AI texture (channel-aware review), and **restore** a human voice (de-slop + voice match + self-audit). Use one layer or all three depending on what the user hands you. ## Pick the mode first - **Generate** — user wants new prose ("write a runbook for X"). Run Construct, then Restore, then a light self-audit. Skip scoring unless asked. - **Review** — user pastes a draft and wants feedback ("does this sound like AI?"). Run Detect → Score → Report. Offer a rewrite. - **Rewrite** — user pastes a draft and wants it fixed ("humanize this"). Run Detect → Rewrite → Self-audit. Show the report only if useful. If unclear, default to **Review** and offer the rewrite at the end. ## Core rules (all modes, non-negotiable) 1. **Never invent specifics to add texture.** No fabricated numbers, names, quotes, dates, or citations. If a draft is vague and a concrete example would help, flag it and leave a `[ADD SPECIFIC EXAMPLE]` placeholder. Sounding human never outranks being correct — this matters most in technical docs. **A specific does not have to contain a digit.** A graded eval found both generate-mode failures were numberless, which is why they got past every check. Treat all of these as fabrication: - **A claim about state.** "Test coverage stayed the same" — invented, in a post whose context file approved exactly one fact. - **A file path, directory or filename....