Text Humanizer Skill

This skill helps users create content that sounds human, not algorithmic. It has three modes:

1. Humanize: Rewrites existing text to remove AI patterns using traffic-light diagnosis, contrastive subtraction, and a staged verification pass.
2. Audit: Diagnoses AI markers in text WITHOUT rewriting. Returns a traffic-light map and pattern list.
3. Generate Prompt: Creates a specialized System Prompt for a specific genre/domain that the user can use in other chats.

Scope boundary — what this is NOT

Long-form fiction is out of scope. A short story, a novel chapter, a screenplay: this skill edits the surface of a text — vocabulary, syntax, punctuation, paragraph rhythm — and in fiction that surface carries a small share of what marks a text as machine-written. Measured: a classifier using only discourse-level narrative features (plot continuity, thematic explicitness, how emotion is conveyed, chronology) separates human from AI fiction at 93.2% macro-F1, and stylistic span-rewriting of the kind this skill performs moved it by 1.6 points, from 95.5% to 93.9% (StoryScope, arXiv:2604.03136; LAMP, CHI 2025).

Running this skill on a story therefore returns a cleaner surface over the same narrative architecture. That is not nothing, and it is not what the request usually means. Say so, and offer the nearest thing that is in scope: --genre blog covers essayistic and personal writing, which is what patterns_creative.md was written for.

Why the skill is not extended to cover it. The architecture does not fit. humanizer.py assembles a prompt from lists of things not to do, filtered by the [A]-[D] tags. A narrative layer is not a list of don'ts: it is a sheet of structural decisions with calibration bands — subplot or none, resolution driven by choice or by circumstance, where the reveal lands — chosen before drafting and applied a few per story. Nothing in that passes through filter_patterns_by_priority. It is a different skill, not a genre in this one.

The description says so too, because that is what routing reads. A boundary stated only here is read after the skill has already been chosen. description used to advertise a "Creative" genre, which is what invited fiction requests; it now says non-fiction and names the exclusion. The genre that "Creative" meant is blog — essayistic, opinion, social and personal writing, the scope of references/patterns_creative.md.

Red Flags (Anti-Rationalization)

STOP and READ THIS if you are thinking:

• "I'll just read the markdown files and rewrite by hand" → WRONG. Run scripts/humanizer.py. It merges universal patterns with genre exceptions, applies the intensity filter and emits the verification steps; done by hand, the genre exceptions are the first thing dropped.
• "The whole text reads like AI, I'll rewrite all of it" → WRONG. Traffic-light diagnosis exists because over-editing introduces AI patterns. Green paragraphs are not touched.
• "More edits means more human" → WRONG. One contrastive substitution beats three stylistic edits. Volume of change is not the metric.
• "The user gave writing samples, I'll imitate the vibe" → WRONG. Build the voice passport along the five documented dimensions and pass it with --voice; an unstructured impression is not reproducible across paragraphs.
• "It's legal/technical text, but max intensity is safer" → WRONG. Intensity is auto-derived from genre for a reason: a D-level stylistic edit in legal text changes what the sentence commits to.

Rationalization Table

Execution Mode

• Mode: hybrid
• Why this mode: scripts/humanizer.py is deterministic assembly — it merges the universal pattern list with the genre list, filters both by the resolved intensity, and injects style, voice passport and extra rules into assets/generator_template.md. The rewriting itself is model judgement performed against the prompt that assembly produces. No script edits text, and a hand-assembled prompt drops the genre exceptions first.

Script Contract

• Command: python3 scripts/humanizer.py --genre {encyclopedic|academic|technical|journalistic|science|blog|social|marketing|corporate|food|crypto} [--style STYLE] [--mode humanize|audit|prompt-gen] [--intensity auto|max|high|medium|low|minimal] [--task TEXT] [--voice PATH] [--extra-rules TEXT]
• Where the input text comes from: not from the script. humanizer.py has no positional argument and no --input; it never reads the user's text. It writes a system prompt, and you apply that prompt to the text held in the conversation. --task is a one-line description of the job, not the text itself.
• Inputs: flags only. --genre is the sole required flag. Defaults: --mode humanize, --intensity auto (resolved per genre), --style falls back to the genre name, --voice and --extra-rules empty.
• Outputs: the assembled prompt on stdout. No file is written and no model is called. Warnings — --style not found, --voice file missing — go to stderr and leave the exit code alone.
• Exit codes: 0 whenever a prompt is produced; 2 for argparse usage errors (missing or unknown --genre, --mode, --intensity). A --voice path that exists but is not a readable file (a directory) raises through as exit 1.
• Idempotency: identical flags produce byte-identical stdout, and there is nothing to overwrite.

Safety Boundaries

• Allowed scope: read-only. The script reads its own references/ and assets/, plus the single file named by --voice. It creates, modifies and deletes nothing.
• --voice is uncontained: any readable path is loaded verbatim into the emitted prompt. Pass a scratch file written for this purpose, not a path the user has not seen.
• Destructive actions: none. "Remove AI patterns" in this skill is prose editing the model performs on text in the conversation — reversible, reviewable, never a file mutation.
• Stop condition: Green paragraphs are not rewritten, and audit mode stops at the diagnosis. Over-editing reintroduces the patterns the skill exists to remove.

Validation Evidence

• Local verification — stdlib only, no venv; paths are relative to skills/text-humanizer/:
  • cd scripts && python3 -m unittest discover -s tests — 56 tests, OK.
  • python3 scripts/humanizer.py --genre marketing --mode prompt-gen | grep -c '^### [0-9]' — 7; --mode audit 8; --mode humanize 9. The digit is load-bearing: it counts the template's own sections, so a heading inside an injected pattern file cannot move the number.
  • --genre technical emits [A] sections only, --genre marketing all four; --genre bogus exits 2, --style bogus exits 0 with a stderr warning.
• Behavioural evals: python3 evals/selftest_evals.py — 83 cases, zero tokens, no agent spawned; it measures the instrument, not the skill. A campaign measures the skill: evals/run_humanize.py then evals/grade_run.py. 21 cases on four axes — failure-mode, coverage over every genre and style file, pressure, and natural prose copied verbatim from repository files. Latest sweep, claude-sonnet-5, 2026-09-04: 75 of 80 checks, 21 of 21 measured. Two comparisons carry an interval: the four paired cases under a neutral brief give delta +0.105, 95% CI [0.028, 0.195], and the six pressure/natural cases give +0.167, CI [0.042, 0.286] — the skill helps more when the brief leans on it, which is what its doctrine is for. Each report records the SHA of the assembled prompt its runs saw and flags itself stale when the skill moves. Trigger evals are a separate check with its own set, evals/trigger/evals.json, run by skill-creator/scripts/run_eval.py. Figures, the campaigns, the self-audit against docs/Manuals/skill-evals_guide.md §11.1, and what a single draw does not license: evals/README.md.
• Gate: python3 skills/skill-enhancer/scripts/analyze_gaps.py skills/text-humanizer from the repo root — no [Execution Policy] advisories.

Usage

Mode 1: Humanize Text

Rewrites text to remove AI patterns with adaptive intensity.

python3 scripts/humanizer.py --genre [encyclopedic|blog|marketing|...] --style [crypto|food|science|...] --mode humanize --intensity [auto|max|high|medium|low|minimal]

Mode 2: Audit Text

Diagnoses AI markers without rewriting. Use when the user wants to see what's wrong before committing to changes.

python3 scripts/humanizer.py --genre [encyclopedic|blog|marketing|...] --mode audit

Mode 3: Generate System Prompt

Creates a reusable prompt.

python3 scripts/humanizer.py --genre [encyclopedic|blog|marketing|...] --style [crypto|food|science|...] --mode prompt-gen --intensity [auto|max|high|medium|low|minimal]

Optional: Voice Passport

When the user provides writing samples, follow this workflow:

1. Read references/voice_passport_template.md to understand the 5 analysis dimensions.
2. Analyze the user's writing samples along those dimensions.
3. Write the resulting 3-5 line voice passport to a temp file (e.g., /tmp/voice_passport.md).
4. Pass the file to the script:

python3 scripts/humanizer.py --genre blog --mode humanize --voice /tmp/voice_passport.md

If no --voice is provided, the default voice is "smart person explaining to a friend over coffee."

[!IMPORTANT] Red Flag: Stop if you think "I'll just read the markdown files manually." WRONG. You MUST run scripts/humanizer.py. It handles the complex logic of merging universal patterns, rewriting strategy, genre-specific exceptions, intensity filtering, and verification steps. Reading files manually introduces human error and laziness.

Key Concepts (v2.0)

Pattern Priorities (A/B/C/D)

Not all patterns are equal. Each is tagged:

• [A] Critical -- Always fix (AI vocabulary, chatbotisms, negative parallelism, em dash abuse), subject to the whitelist in references/patterns_universal.md: a habit evidenced in the voice passport, a domain term you can point at, and quoted material are not findings. [A] is the only class that fires at low and minimal, so it is the class where a false positive costs most.
• [B] High -- Fix in all modes except legal text.
• [C] Medium -- Fix in full editing and expert content.
• [D] Stylistic -- Fix by context (Rule of Three, Synonym Cycling, Colon Disease).

Intensity Levels

Controls how many priority levels to fix. Auto-detected from genre:

• max (marketing, social) -> A+B+C+D
• high (blog, food, crypto) -> A+B+C
• medium (corporate, journalistic, encyclopedic) -> A+B
• low (technical) -> A only
• minimal (legal) -> A only, cautiously

Traffic-Light Diagnosis

Before rewriting, paragraphs are classified:

• Red (3+ markers) -> Full rewrite.
• Yellow (1-2 markers) -> Spot fix.
• Green (clean) -> DO NOT TOUCH. Over-editing introduces new AI patterns.

Contrastive Subtraction (CoPA)

Beyond removing bad patterns, actively replace the most predictable word in each sentence with a less probable but appropriate alternative. One such replacement beats three stylistic edits.

Verification Passes

After rewriting: (1) scan for leftover patterns; (2) read as a stranger, then ask the opposite question and report over-correction separately -- an overshooting edit leaves its own fingerprint, and the target is the human band rather than the pole opposite the AI one; (3) check sentence length variance ("cardiogram" for 300+ word texts), bounded by a ceiling so spikes stop at the band; (4) at max and high intensity only, the outline test -- read the first sentence of each paragraph as a list, and if it forms a clean summary the structure is machine-shaped. Pass 4 is stripped from the emitted prompt at every lower intensity, because a clean outline is the goal in an API reference, a news lead and a BLUF memo.

Resources

• Taxonomy & Intensity
• Universal Patterns (The "Don'ts")
• Rewriting Strategy (The "How")
• Encyclopedic Patterns
• Creative Patterns
• Voice Passport Template
• Domain Styles
• Prompt Template
• Behavioural evals — twenty-one cases on four axes, deterministic grading, committed baseline