Humanizer

Humanizer

A rewriting engine that makes text sound like it was written by an intelligent, specific human — not a language model — while never touching the underlying facts, claims, or accomplishments.

This is not a detector-evasion tool. The target is genuine, voice-consistent, specific prose. If a rewrite ever starts to feel like it's performing "humanness" (forced typos, injected slang, random incorrect grammar) rather than just being clear and specific, that's a sign to stop and reconsider — see references/falsely-accused-patterns.md.

Non-negotiable rules (never violate these for style reasons)

1. Preserve every idea, fact, number, name, and claim exactly. Never round up, generalize, or soften a real number; never inflate a scope or accomplishment.
2. Never invent experiences, quotes, statistics, or details that weren't in the source text or provided by the user.
3. Never change what actually happened — accomplishments, placements, dates, partnerships stay exactly as stated (or exactly as scoped-down if the user indicates uncertainty — see EverGreen precedent in references/voice-profile.md).
4. Never add fluff, motivational flourishes, or claims to "sound better." If cutting a forbidden pattern would leave a gap, fill it with a specific real detail from the source — never with generic filler.
5. Preserve professionalism and technical precision appropriate to context — see references/register-guidance.md.
6. When genuinely unsure whether a claim is confirmed, ask the user rather than guessing at how to phrase it.

Workflow

1. Identify the register. Check references/register-guidance.md for the specific context (college essay, professional email, LinkedIn, technical docs, startup pitch, grant, research writing, presentation, website copy, casual message). This determines which patterns matter most and what "good" looks like here.
2. Read the draft (or generate a first draft) normally. Don't try to write "humanized" from scratch — draft for content and accuracy first.
3. Pass 1 — Forbidden pattern scan. Check the draft against references/forbidden-patterns.md. Flag every instance of: repetitive cadence, filler transitions, hedge-stacking, the AI-word list, buzzword/cliché density, overexplaining, generic conclusions, fake enthusiasm, rhetorical-question crutches, adjective/adverb stacking, uniform paragraph rhythm.
4. Pass 1.5 — Don't over-correct. Cross-check every flagged item against references/falsely-accused-patterns.md before removing it. A single em-dash, one instance of "delve" in a fitting context, correct semicolons, and earned parallelism are not violations — leave them.
5. Pass 2 — Rewrite toward positive principles. Apply references/positive-principles.md: vary sentence length deliberately (burstiness), put the most important information at the end of sentences (stress position), replace abstractions with the specific real detail already in the source, let paragraph length track the actual weight of the idea rather than a template.
6. Pass 3 — Voice match. Apply references/voice-profile.md if rewriting for R-sign specifically: match the correct register (raw brainstorm / clean explainer / polished outreach), keep technical numbers and names exact, keep directness, don't add inspirational language that isn't earned by the content.
7. Pass 4 — Final quality check. Run the checklist below before returning the result.
8. Deliver the rewrite along with a brief, plain note on what changed and why (not a lecture — one or two lines, e.g., "Cut the 'furthermore' chain and varied sentence length in paragraph 2; kept your numbers exactly as given.").

Quality checklist (run before returning any rewrite)

• Every number, name, date, placement, and claim matches the source exactly — nothing invented, inflated, or softened without the user's indication.
• No more than one item from the AI-word list (§5.2 in forbidden-patterns.md) survives, and only where it's genuinely the best word for that register.
• No stacked transition words (furthermore/moreover/additionally) — at most one heavier transition per document, used only when it earns its place.
• Sentence lengths in any 3-sentence stretch are not all within a few words of each other — real variance is present.
• No generic "In conclusion" / restate-everything ending, unless the register (long technical/academic doc) genuinely calls for a summary.
• No hedge-stacking on claims the user is actually confident about; no false confidence on claims that are genuinely uncertain.
• No rule-of-three lists that were forced to fit three when the real count is different.
• Register matches context (see register-guidance.md) — a casual message wasn't over-formalized; a grant application wasn't under-formalized.
• If this is for R-sign: technical precision preserved, directness preserved, no added inspirational flourish, correct register (brainstorm/explainer/outreach) used.
• Read it once out loud (mentally) — does it sound like one specific person said this, or like a template with the blanks filled in?

Reference files

• references/forbidden-patterns.md — full "Do Not Use" database: every flagged pattern with why/example/when-acceptable/when-avoid/alternatives.
• references/falsely-accused-patterns.md — patterns that are NOT actually AI tells; don't strip these on sight.
• references/positive-principles.md — evidence-based traits of genuine human prose to write toward.
• references/register-guidance.md — what to hunt for and what "good" looks like in each of the 12 covered contexts.
• references/voice-profile.md — R-sign's specific observed voice, registers, and defaults.

Configuration / extensibility

• Strictness: default is "moderate" — flag and fix clear violations, don't nitpick borderline cases. If the user says "be aggressive" or "light touch," adjust pass 1 accordingly.
• New patterns: if the user identifies a new tell they notice in their own writing or in Claude's output, add it to forbidden-patterns.md following the same why/example/when-acceptable/when-avoid/alternatives structure, so the database keeps growing.
• New registers: if a new context comes up repeatedly, add a section to register-guidance.md.
• Voice drift: if the user's writing style evolves (new job, new register), update voice-profile.md rather than letting the skill rewrite toward a stale profile.
• Detector-evasion requests: if a request drifts toward "make this pass an AI detector" rather than "make this sound like me," redirect toward genuine voice-matching — the two usually produce the same result, but if they conflict, prioritize genuine voice.