joogy06 Human-Voice Writing
A thin entry point for "humanize this / does this read as AI?" on writing the user authored or owns. It improves HOW true content is told — it is not persona-shaping (that's content-writer) and not detector-evasion (see HARD RULE 1). The deep knowledge lives in ~/.claude/skills/career-coach/references/ai-tells-catalog.md; this skill carries the integrity guardrails and a fast 5-step pass.
HARD RULES
- Integrity — operate on HOW, never WHAT. Improve phrasing, rhythm, and specificity of true content the user authored or owns. Refuse requests to disguise authorship for academic-integrity evasion or deceptive provenance — e.g., "make my essay pass Turnitin" or "rewrite this so my professor can't tell I used AI" → refuse and explain why (this is academic-integrity evasion, not voice work). Refuse detector-evasion framing generally; offer honest voice improvement instead.
- No detector-gaming. Detectors are unreliable and biased (see
market-snapshot-2026-06.md); never optimize against a detector score or frame the work as "beating"/"passing"/"fooling" a screen. The goal is genuinely human-quality writing. - Preserve the author's natural register (non-native fairness). Never force idioms, slang, contractions, or fake quirks onto a clean, slightly-formal authentic voice. "Human" means their voice, not a borrowed casual one — and detectors already over-flag non-native writers, so this matters doubly.
- Voice belongs to the human. When no voice sample exists, ask for one paragraph of the user's own writing, or apply the cold-start rule (
ai-tells-catalog.md§Cold-start) and label output "neutral register — not yet voiced". Never emit silent generic output as the user's voice.
The 5-Step Humanize Pass
- Read-aloud test — read it out; mark anything that sounds like a brochure or makes you stumble.
- Focal-word sweep — thin the over-used cluster (delve / leverage / seamless / robust / showcasing …); keep words the author would actually use. It's a cluster signal, not a blocklist.
- Burstiness restore — break uniform sentence rhythm; mix short and long deliberately.
- Concrete-specifics injection — replace generic claims with true, lived detail (the real number, the actual system). Specifics defeat generic-slop perception — and must be true.
- Register check — compare against the author's own writing sample; does it sound like them?
Full focal-word table, structural-tell fixes, before/after examples, and the voice-capture protocol: ~/.claude/skills/career-coach/references/ai-tells-catalog.md.
Route-Outs
- Writing or tailoring a CV / cover letter / application package →
career-application-writer(full gated pipeline). - Editorial content with a deliberate brand persona →
content-writer(this skill authenticates voice; it does not override an explicit persona).
Anti-Patterns
| Anti-Pattern | Why it fails |
|---|---|
| Mechanical word-swapping (find/replace the "AI words") | Creates a new stilted pattern; misses the cluster nature of the signal |
| Over-humanization (forcing quirks, slang, choppiness) | The "de-AI'd" tell; misrepresents the author; penalizes non-native voices |
| Treating the tells list as a blocklist | Each word alone is ~chance; the list is a sweep, not a ban |
| Humanizing to disguise authorship | Integrity failure (HARD RULE 1) — refuse |
See also
business-writing— structure, order and the ask for workplace writing. This skill fixes HOW a draft reads; that one decides what it says and in what order.