Humanizer

Humanizer

Remove artificial patterns, synthetic cadence, and formulaic tells from user-facing prose. Precision beats style. This is an editorial pass for writing intended for human readers.

Boundaries

• Never auto-humanize ordinary coding output, diffs, PR summaries, or CLI tool output.
• Never rewrite source code, tests, CLI commands, YAML frontmatter, citations, or locked SmartDoc content.
• Never invent facts, quotes, statistics, or sources to make text feel more colorful.
• Voice sample provided by the user overrides default pattern recommendations.
• Academic manuscripts after GOAL_LOCK route to academic first; humanizer is an optional final prose polish only when requested.
• Technical documentation structure (Diátaxis / STE) routes to /technical-writing.

Semantic-Preservation Constitution

These tokens must survive byte-equivalent unless the user explicitly requested changes:

• Facts, numbers, dates, personal names, proper nouns
• URLs, citations, file paths, line references
• Code snippets, CLI commands, flags, API endpoints, error strings
• Direct quotes
• Negations (not, never, unless, without)
• Security, compliance, and legal wording
• Legitimate domain terms when they represent the actual technical concept (interface, vector, primitive, surface, modality, schema)

Replace an abstract term only when it is empty marketing jargon (e.g. marketing "vector" → "pathway"; linear algebra vector stays untouched).

The Five Pattern Families

Audit text against the five families of AI writing tells (detailed catalog in references/patterns.md):

1. Staging: Theatrical setups, announcements, throat-clearing, and artificial transitions ("In today's fast-paced world", "Let's delve into", "It is important to remember"). Cut the preamble; start with the substance.
2. Rhythm: Monotone sentence lengths and predictable cadence. AI drafts often output uniform 15-20 word sentences. Vary sentence length dynamically: combine punchy fragments with nuanced multi-clause observations.
3. Inflation: Hyperbolic puffery, superlatives, and buzzword padding ("pivotal", "testament to", "game-changing", "seamlessly", "foster", "garner", "tapestry"). Use grounded, factual vocabulary.
4. Formatting: Obsessive bullet lists where prose fits better, excessive bolding on lead-in words, and overused em dashes (—). Convert mechanical lists to flowing paragraphs; replace em dashes with commas or separate sentences.
5. Leftovers: Chatbot residue and formulaic conclusions ("Certainly!", "I hope this helps!", "In conclusion, the future looks bright"). Remove conversational fluff and generic wrap-ups.

Workflow

1. Ingestion Mode

• Paste Mode: User provided an excerpt directly. Edit and output the rewritten prose with an optional summary of removed patterns.
• File Mode: Target file specified. Read the target, inspect surrounding context, edit user-facing text, and preserve Markdown formatting and code blocks unchanged.

2. Analysis Pass

• Scan for the five pattern families.
• Identify the author's intended tone (casual, authoritative, editorial, journalistic).
• Note any user voice sample provided to calibrate rhythm and vocabulary.

3. Rewrite & Refine

• Strip throat-clearing and puffery.
• Break monotonous sentences and vary paragraph lengths.
• Replace AI vocabulary with direct, everyday verbs and nouns.
• Keep domain terminology intact.

4. Integrity Check

• Audit against the Semantic-Preservation Constitution: ensure no dates, numbers, URLs, or technical terms were altered or dropped.
• Self-audit: "Does any sentence still sound generated by a committee or an LLM?" If yes, tighten further.