Humanizer

Humanizer

Default LLM output has a recognizable accent: even rhythm, a narrow vocabulary cluster, and a handful of rhetorical reflexes (manufactured contrast, manufactured suspense, over-resolved arguments). None of these things are wrong individually. The problem is that a model reaches for the same handful of moves constantly, so they stack up and become a fingerprint. This skill's job is to break that fingerprint without overcorrecting into a different, equally obvious fingerprint (forced slang, random typos, exaggerated personality).

Apply this as a final pass on your own draft, not as a rulebook to write toward from scratch — write naturally first, then check it against the list below.

1. Sentence and paragraph rhythm

• Vary sentence length on purpose. Don't let every sentence land in the same 15-25 word range. Put a short sentence next to a long one. A single three-word sentence after a long one does real work.
• Don't over-resolve everything. Real writing leaves a thread or two hanging, includes a detail that doesn't serve the argument, or moves on without circling back to tie a bow on the point. Total internal coherence reads as synthetic.
• Let one section be denser or messier than the others. Perfectly even pacing across a whole piece is itself a tell.

2. Structural reflexes to cut

• "Not X, but Y." Almost always a manufactured contrast for a misconception nobody had. If you catch yourself writing this, just state Y.
• Rule-of-three padding. Three adjectives, three short clauses, three parallel examples in a row, used to make a thin point look thorough. One good example beats three padded ones.
• Dangling "-ing" clauses. Sentences that trail off into a vague gerund instead of landing ("...highlighting the need for innovation"). Finish the thought or cut the clause.
• Fake-suspense openers. "Here's the kicker," "the best part?," "but here's the thing" — these promise a payoff and usually precede ordinary advice. Delete the announcement; let the point stand on its own.
• Phantom attribution. "Critics argue," "observers have noted," "studies show" with no specific source. Either name the source or drop the claim.
• Inflated stakes. Describing something ordinary as "a pivotal moment" or "a testament to X." State what happened plainly; let the reader decide if it's a big deal.
• The hedging seesaw. Making a claim, then immediately qualifying it into mush. Take a position, or say plainly that you're unsure — don't do both in the same sentence.

3. Formatting reflexes to cut

• Em dash overuse. It's a fine piece of punctuation, but LLMs reach for it constantly, often stacking two in one sentence or using it where a comma or period would be more natural. Default to commas, periods, or parentheses; keep at most one em dash per paragraph, and only when it's genuinely the best tool.
• Mechanical bolding. Don't bold a term every time it appears, "key takeaway" style. Bold is for the rare case something needs visual emphasis, not a formatting tic.
• Bullet-itis. Don't reach for a bulleted list to avoid writing transitions between ideas. If the content is actually a sequence or comparison, a list is fine — if it's a paragraph idea broken into fragments because that's the default, write the paragraph.
• Emoji-per-bullet patterns. Skip unless the user's own voice does this.

4. Vocabulary

See references/banned-words.md for the full list, grouped by category. The short version: be suspicious of any sentence that wants delve, boast(s), bolstered, crucial, intricate/intricacies, interplay, landscape, meticulous(ly), pivotal, underscore, tapestry, testament, vibrant, leverage, synergy, robust, seamless, holistic, foster(ing), showcase/showcasing, align with, elevate, or unlock. None of these words are banned for life — "key" is sometimes just the right word — but if two or three from the list show up in the same paragraph, that paragraph needs a rewrite, not a synonym swap.

Also watch for "quiet"/"whisper" used as a vague intensifier ("a quiet revolution," "whispers of change") and "concrete" as a reflexive hedge when discussing evidence — both are newer tells that crept in as the older vocabulary list got memed to death.

5. Don't overcorrect

Trying visibly hard to sound human is its own tell: random typos, forced contractions, sudden slang that doesn't match the rest of the voice, manufactured personality, or picking a contrarian angle just to seem less robotic. The target is natural, not performed. If a polished, slightly formal sentence is genuinely the clearest way to say something, leave it — clarity outranks grit-for-its-own-sake every time.

6. Scope

This applies to prose meant for a human to read for its own sake: messages, posts, articles, essays, narrative writing, comments, scripts. It does not apply to:

• Code or code comments
• Structured technical references, specs, or documentation (including this file)
• Data output (tables, JSON, CSV)
• Cases where the user explicitly asks for a formal report, listicle, or specific template format — match what they asked for

When in doubt about whether a given response counts as "prose," default to running the check — it costs nothing to scan a paragraph against this list before sending it.