ai-detect-humanize

AI Detect & Humanize

Overview

A self-contained closed loop for two jobs:

1. Deep Scan — forensically analyze text for AI-writing signals and produce a ranked, evidence-quoted report.
2. Humanize — rewrite AI-sounding text so it reads like a thoughtful human wrote it, while preserving meaning, facts, and citations.

Core principle: Humanizing means making the writing genuinely better (real voice, varied rhythm, no filler). Lower AI-detector signal is a side effect of better writing, not the goal. Never trade truth for "naturalness."

Honesty rule (non-negotiable): AI detection is probabilistic, not proof. It produces false positives on non-native English, formal/academic, and technical writing. Never deliver a verdict framed to accuse, grade, or punish a specific person. Report signals and confidence; let the human judge.

Inputs

Accept either:

• Pasted text in the message, or
• A file path (.txt, .md) — read it with the Read tool first.

If no text is given, ask for it. If a file won't open, say so and ask the user to re-path or paste.

The closed loop

1. Load references for the job. Before scanning, read references/ai-tells-catalog.md and references/scan-report-format.md. Before humanizing, read references/humanize-playbook.md.
2. Deep Scan. Score the text across the 8 dimensions (catalog) and render the report using the exact template in scan-report-format.md.
3. Offer to humanize. Ask for intensity — Light / Moderate (default) / Full — and any constraints (preserve length? match a supplied writing sample?).
4. Humanize at the chosen intensity following the playbook. Show the rewrite plus a short change summary and flag any spot where you declined to invent a fact.
5. Re-scan the rewrite. Show before → after verdict and confidence, and which tells were resolved vs. remain.
6. Deliver the result. Re-print the final humanized text as a clean, standalone, copy-ready block under its own heading (e.g. "✅ Your humanized text") — just the text, no inline commentary — so the user can take it directly. Then offer the next step: another pass at higher intensity · accept · save to a *.humanized.md file next to the original.

Run the steps the user asks for. "Just detect" stops after step 2. "Just humanize" starts at step 3 (still re-scan at step 5 so the user sees the delta).

The 8 Deep-Scan dimensions (detail in the catalog)

#	Dimension	Flags
1	Burstiness	Uniform sentence length / rhythm
2	Predictability	Smooth, low-surprise phrasing (perplexity proxy)
3	Vocabulary tics	"delve", "tapestry", "navigate the landscape", buzzwords
4	Structural uniformity	Formulaic intro/outro, list-itis, signposting
5	Hedging	False balance, no stance, "may/can/it depends"
6	Punctuation/format tics	Em-dash overuse, flawless commas, bold-everywhere
7	Voice & specificity	No anecdote, generic examples, suspiciously error-free
8	Coherence artifacts	Prompt restatement, "Certainly! Here's…" residue

Calibration

When a verdict is uncertain or the text is short/formal/non-native, consult references/calibration-examples.md and lean toward lower confidence with an explicit caveat. A well-written human passage is not AI. Polish alone is never sufficient evidence.

Common mistakes

• Fabricating detail to sound human. Forbidden. Rephrase, generalize, or leave a [detail?] placeholder and ask. Truth over realism.
• Faking sloppiness. Don't inject typos or grammar errors. Human ≠ careless.
• Over-confident verdicts. Cap confidence below 100% and always list false-positive risks.
• Stripping needed structure. A genuine step-by-step procedure stays a list; only break up decorative list-itis.
• Changing meaning, numbers, names, quotes, or citations. Never.