Audit any draft against the full anti-AI-writing canon, then rewrite the failing parts so the output reads like a sharp human typed it. Three-layer detection: rule packs (Vale) → semantic check → LLM rewrite. Single-pass operator workflow.
Source synthesis: Wikipedia WikiProject AI Cleanup field guide (rule-source-of-truth), blader/humanizer v2.5.1 (workflow + audit pattern), Cole Schaefer (Honey Copy sentence craft), Joanna Wiebe (Copyhackers conversion canon), per-claim factuality eval (Hamel + Shreya — Evals are systematic data analysis on your LLM application — start with error analysis, not tests, Build LLM-as-judge as binary true/false, one judge per pesky failure mode — and validate against human labels).
When to use
Any customer-facing draft before publish: LP, email, blog, content, ad, video script. Especially before any LLM-drafted content that the team will sign their name to.
The 11 detection rule packs
| Rule pack | Severity | Source |
|---|---|---|
| KillList | error | operator-voice canon |
| WeakVerbs | warning | sentence craft |
| AbstractNouns | warning | conversion canon |
| ThroatClearing | error | operator-voice |
| Cliches | warning | AI-fingerprint canon |
| BannedVocabulary | error | brand voice DNA — dead AI vocabulary |
| NegativeParallelism | error (the big one) | the most common AI tell — "It's not X, it's Y" |
| MetaCommentary | error | "In conclusion," "It's worth noting that…" |
| ParticiplePhrases | warning | overused construction in AI prose |
| ChatbotArtifacts | error | "Certainly!", "Great question!", "I'd be happy to…" |
| CopulaAvoidance | warning | overuse of "to be" verbs |
| DeadPhrases | error | "delve," "unleash," "tapestry," "navigate the landscape" |
Negative parallelism is THE BIG ONE because every LLM produces it dozens of times per response. Detection alone catches the gap; humanizer closes it with a rewrite layer.
Three layers
1. Detection. Vale rule packs flag every violation with line, match, severity, and fix hint.
2. Semantic check. LLM ingests surrounding paragraph + Vale finding + voice DNA principles → produces a single replacement sentence/clause that fixes the violation, preserves the underlying claim, matches the asset's voice. Non-destructive: writes to a sibling file for human review.
3. Audit signature pass. After detection + optional rewrite, ask the LLM the blader/humanizer audit question:
> "What makes this obviously AI generated?"
Record the answer. If it names new tells the rule packs missed, add them to the rule-pack queue. This is the auto-improvement loop.
Output
- Per-violation table: rule × severity × line × match × why × fix-hint.
- Single-line summary: PASS / FAIL N errors M warnings.
- Sibling file
.humanized.md
Exit codes
- 0 — clean (warnings allowed unless strict).
- 1 — error-level violations found.
- 2 — file/tool not available.
Composes with
- Voice-enforce gate — voice-enforce is the gate; humanizer is the editor that produces the rewrite.
- CRO rubric — distinct surface. CRO fails on structure; humanizer fails on voice.
- Preflight — humanizer is the strict superset of voice-enforce.
Refusal patterns
- Substrate gap. If brand-voice canon is missing or stale, refuse to rewrite — would invent voice instead of inheriting it.
- Internal-spec asset (substrate, raw, intel) — bypass automatically.
- Empty asset / non-markdown — exit 2 with explanation.
Calibration
Track under taste-type "voice-enforcement" and "humanizer-rewrite." Brier signal: post-publish reader-detection rate. Look for low AI-detection rate on the published page.
Common failure modes
- Rewriting without the brand-voice canon loaded → drift. The rewrite invents voice instead of inheriting it.
- Treating the rule packs as static. Every new model variant produces new tells; the audit-layer feedback loop must run.
- Auto-merging rewrites without human review.
- Running humanizer AFTER claim-verify. Run voice gates before truth gates so substrate-cited claims are not over-rewritten.
- Stopping at detection. Detection without rewrite leaves operators to fix by hand.
Why this matters
Detection alone catches the gap. Without the rewrite layer, operators get a fail report and fix by hand — that is the gate's role, not the editor's. With the humanizer, the operator gets a draft + a reviewable rewrite in one pass. The compound is read-time-to-ship reduction.