anti-ai-writing-humanizer

Write prose a human can deliver as their own, or rewrite existing prose so it stops reading as AI. Same rules either way — the only thing that changes is how deep to go.

Pick the depth (the rules are the same; the thoroughness isn't)

Default to light for a short snippet, heavy for a whole document or an explicit deep request. If genuinely unsure, ask once: "quick pass or deep audit?"

This is for prose — emails, posts, drafts, chat replies, and technical prose too (READMEs, docs, issue write-ups), where §8 of write-human.md does most of the work. Never run it on code, code comments, or technical config.

Sources of truth (and who wins)

1. House rules (always win): references/house-rules.md — the editable list of banned

vocabulary, banned structures, the zero em-dash rule, straight quotes, sentence-case and contraction defaults. Read it every pass. It's the single source of truth; on any conflict with the engine, the house file decides. This is the one file you customize to your own taste.

1. The prescriptive ruleset: references/write-human.md — the core "write this way, not that

way" rules, synthesized from the house rules, the Wikipedia "Signs of AI writing" catalog, the engine, and the plain-speech ruleset (§8). The main guide for both depths and for writing from scratch.

1. The deep engine (heavy only): references/avoid-ai-writing.md — the conorbronsdon

avoid-ai-writing engine (v3.9.0, MIT). ~50 pattern categories, tiered vocab tables, P0/P1/P2 severity, context/voice profiles, stylometric rhythm tests. The detection superset for audits.

When writing from scratch

Apply the rules as you write rather than fixing afterward — clean-first keeps the voice consistent (stitched-together rewrites cause the "sudden style shift" tell themselves). Lead with the specific fact or ask, use plain words, vary sentence length, and where you lack a specific (a number, name, date) leave a clearly marked blank for the user instead of padding or speculating.

When rewriting (the pass)

Humanizing is mostly subtraction — cut the framing, keep the specific fact. Don't invent "human texture"; that's its own tell.

1. Plain words. Replace inflated vocab (delve, showcase, leverage, foster, garner, crucial,

tapestry, landscape, pivotal, vibrant, "serves as", "valuable insights"...) with the ordinary word, or cut it. Never swap one banned word for another ("leverage" → "harness" is still a tell).

1. Direct sentences. Kill "not just X, but Y" parallelism; the "Despite its X… continues to

thrive" formula; rule-of-three padding; synonym chains (repeat the noun). Use is/are/has.

1. No AI framing. Cut significance-claiming, collaborative voice ("Let's explore", "As we can

see"), knowledge-cutoff disclaimers, and verbalized placeholders.

1. End at the fact. Remove "-ing" tack-on clauses and "challenges and future outlook" kickers.
2. Quiet punctuation and formatting. Sentence case headings, bold only for real emphasis,

contractions in chat, no emoji unless the user used them. No em dashes at all — and don't swap in an en dash, a spaced hyphen, or parentheses doing the same job; end the sentence or use a comma. No colon splicing two clauses. Straight quotes only.

1. Plain speech (`write-human.md` §8). Name the mechanism, not the mood: cut abstract metaphor

nouns (substrate, vector, surface, flywheel), mannered flourishes, and adverbs propping up weak verbs; make passives name their actor; split sentences the reader has to re-parse, but don't clip them into symbol-speak either. Apply the portability test — a sentence that would fit unchanged in someone else's document says nothing, so replace it with a fact or cut it.

1. (heavy) Run the engine across all categories, pick a context/voice profile if the piece has

one, fix P0+P1 always and P2 when warranted, apply house rules as the overlay, then re-scan and fix again until a pass produces no new edits.

Surgical only. If a sentence already reads human, leave it. Over-polishing clean prose is its own tell. Preserve meaning, facts, numbers, names, quotes, and citations exactly.

Report back

Depth: light | heavy   (heavy: profile blog, voice blunt)
- Replaced: leverage→use, delve→look at, "serves as a testament"→shows  (N)
- Cut: 2 rule-of-three, 1 optimism kicker, 1 template outline reshaped
- Plain speech: 2 passives given actors, "substrate"→"base", 1 unportable sentence cut
- Rhythm (heavy): 4 short sentences added to break uniform pacing
- Em dashes: 6 → 0   Colon splices: 2 → 0   Curly quotes: straightened
- Passes to convergence (heavy): 2

If it's already clean, say "No changes — already reads human." Never invent edits to look busy.

Updating this skill (one owner per rule — edit in one place)

To avoid the same rule drifting across files, each kind of change has a single home:

Credits

Built from the Wikipedia "Signs of AI writing" catalog and Conor Bronsdon's avoid-ai-writing engine (MIT, vendored at references/avoid-ai-writing.md), synthesized into references/write-human.md. See README.md and NOTICE for full attribution.