Humanizer

Humanizer

Rewrite text so it stops reading as machine-generated. The real failure mode of AI prose is regularity, false specificity, and ceremonial filler — not any single banned word. Fix the underlying patterns, don't just swap vocabulary.

Based on Anbeeld's WRITING.md (https://github.com/Anbeeld/WRITING.md), MIT-licensed. The references in this skill adapt that ruleset for English and German.

Important framing

Two failure modes that pull in opposite directions:

1. AI texture — generic anchors, parallel three-part lists, "is a testament", "plays a pivotal role", "experts say", every paragraph the same shape, em dashes everywhere as the default connective.
2. Fake-human overcorrection — invented typos, forced slang, staged uncertainty, manufactured negativity, programmatic sentence-length wobble. Just as obvious, often worse.

The skill avoids both. The fix is not "add soul" via tics — it is concrete anchors, plain words, broken regularity, real stance where the genre invites it, and neutrality where the genre expects it. Fit the context.

Workflow

1. Detect language

English vs German from the input. If mixed (e.g., German body, English headlines), humanize each section with its own pattern set. If unsure or the input is too short to tell, ask once.

2. Read the language reference

• English → references/english-patterns.md
• German → references/german-patterns.md

The references hold the actual rules, watch-lists, and before/after examples. Don't humanize from memory — load the file. The German file has German-specific AI tells (spielt eine zentrale Rolle, Insgesamt zeigt sich, nominalization stacks, „..." typography, etc.) that the English file does not cover.

3. Anchor before rewriting

Before changing words, decide:

• Medium and audience — chat reply? landing page? legal disclaimer? founder's blog post? marketing email? Format and stance are part of register.
• What the reader needs first — the answer, a next action, a verdict, a story?
• Whether the genre expects a visible writer — opinion, review, retrospective: yes. Documentation, neutral summary, news-style: no. Don't manufacture stance where none belongs, don't sand stance off where it does.

If the user said "keep it formal" or "this is a legal notice", do not inject voice. Humanizing legal copy means cutting filler and false certainty, not adding personality.

4. First pass — cleanup

Run the patterns from the language reference. Targets:

• Inflated significance and ceremonial phrasing (stands as a testament, plays a pivotal role)
• Promotional / brochure tone where it doesn't belong
• Vague authority laundering (experts say, Studien zeigen) without a named source
• Negative parallelism (not just X, but Y), reflexive rule-of-three, elegant variation
• Generic anchors (many, various, several, meaningful changes)
• Hidden mechanism claims and unsupported causality (drove, proved)
• Paragraph-shape regularity (every paragraph: claim → orderly elaboration)
• Filler openers (It's important to note, It's worth noting)
• Generic conclusions (Overall, Insgesamt zeigt sich)
• Sycophantic openers and canned closers (Great question, I hope this helps)
• Em dash overuse as default connective (the dashes themselves aren't the problem — using one in every paragraph is)
• Curly quotes / single-character ellipses where the medium is plain text
• Knowledge-cutoff disclaimers, emoji decoration, title-case headings

Preserve facts, names, numbers, and quotes. If the source is vague, keep it vague honestly — don't invent a McKinsey report to replace "experts say". Either name a real source the user can verify, or narrow the claim ("some readers report") and flag it.

5. Second pass — concrete anchors and broken regularity

After cleanup, ask:

1. Anchor audit — does each substantial paragraph carry one concrete anchor (proper noun, specific number that isn't just a date, direct quote, named decision, checkable detail)? If a paragraph's most concrete thing is a name and a date, it's still too generic.
2. Regularity tripwire — what's the single most repeated visible pattern in the piece? If the same move appears 3+ times, rewrite at least one. Three-item lists count even when phrased as X, Y, and Z.
3. Stance fit — if the genre expects a writer (opinion, review, blog), is the writer visible in one sentence? If the genre expects neutrality (docs, news), is it neutral?
4. Specificity earned — for any added detail, can you actually verify it? If not, cut it. Don't replace generic with fabricated-specific.

Don't add "soul" via fake tics. Add stance only where the genre invites it, and prove it through concrete observation rather than performed personality.

6. Final read-through

Read as a first-time reader. Cut anything that's auditioning. Cut sentences whose only job is to announce the next sentence. Most edits should make the text shorter.

For pieces under ~150 words, steps 4 and 5 are enough. For longer pieces, also check: is the spine catalog prose ("first X, then Y, then Z") or system-tour prose (one paragraph per labeled bucket)? If yes, restructure around one through-line.

Output format

Default structure when the user asks for a humanize pass:

## Cleaned draft
<first pass — generic, ceremonial, vague-authority, regularity removed>

## What still reads AI
- <bullet — the most repeated pattern that survived>
- <bullet — any remaining false specificity, hidden mechanism, generic anchor>

## Final
<second pass — anchors added or vagueness honestly kept; regularity broken>

For very short inputs (one sentence, a tweet, a headline), skip the section headers and give: cleaned → final.

If the user says "just give me the rewrite", return one final version and skip the diagnostic sections.

If asked to also explain why changes were made, add a short rationale per major change — not per word.

German guardrails

Real umlauts only — ä ö ü ß Ä Ö Ü. Never ae oe ue ss. Watch for transliterations like fuer or Massnahme and correct them to für / Maßnahme.

Use German typographic quotes „..." in typeset prose; use straight quotes in plain-text contexts (chat, comments, code). Match the source.

What not to do

• Don't strip facts to make text "flow better". Numbers, names, quotes, dates stay.
• Don't invent specifics to replace vague ones. If the source says "experts argue", don't manufacture a 2023 study. Either narrow the claim or keep it honestly vague.
• Don't inject fake-human moves: typos, slang, forced asides, programmatic short-then-long sentence wobble, manufactured uncertainty (I think... maybe... sort of) when the uncertainty isn't real. These are detector-bait and read as performative.
• Don't flatten useful structure. If the source is a help-center article with descriptive headings, lists for steps, descriptive link text — keep them. Removing structure is not a humanization win for medium where structure helps the reader.
• Don't apply a single voice. Marketing copy stays selling (less floridly). A research summary stays measured (less padded). A founder's blog post can carry stance; a privacy policy cannot.
• Don't exceed source length unless asked. Humanizing usually shortens.
• Don't promise detector safety. The skill makes prose less pattern-rich; it doesn't certify authorship. Surface style is fragile evidence.