Natural Writing

Natural writing

Produce prose that avoids the documented signs of AI writing, based on Wikipedia:Signs of AI writing. That page is a field guide for detecting AI text; this skill inverts it into rules for not producing those patterns.

One pattern alone can be fine. The goal is prose where the patterns don't cluster and nothing reads generated.

Scope: this skill is a guideline, not law. Skimmability for the reader outranks any style rule in it (short front-loaded paragraphs, no text blobs, examples first); accuracy and reference-integrity rules always bind. Under that house standard, bold-label bullets and sparing keyword bolds are sanctioned formatting: the bold and bullet cautions in this skill apply to prose written for external readers, not to this repo's own skill docs.

When to apply

• "humanize this text" / "make this sound less like AI" / "remove the AI tells"
• "does this read as AI-generated?"
• Drafting or rewriting any prose a human will read: documentation, READMEs, articles, blog posts, reports, emails, announcements, PR descriptions
• Any writing task where the user cares about tone, voice, or naturalness

Source of truth

The catalog of signs changes over time (models change; the community adds new tells). Prefer the live page, fall back to the bundled snapshot:

1. Live: fetch https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing (raw wikitext: https://en.wikipedia.org/w/index.php?title=Wikipedia:Signs_of_AI_writing&action=raw). Do this when auditing text for AI tells, when the user asks for a thorough pass, or when the bundled snapshot looks stale for the topic at hand.
2. Snapshot: references/signs-catalog.md is a distilled copy of the full page, reorganized as writing rules. Use it as the working checklist for every invocation; it is enough when offline or for routine drafting.

Snapshot last synced: 2026-08-26, from page revision 1371415133 (2026-08-26T08:17:42Z). See "Maintaining the snapshot" below for how to update it.

references/words-to-avoid.md lists the words and phrases the page flags, organized for scanning, plus ready-made grep commands for the mechanical checks.

Workflow

1. Load the rules

Read references/signs-catalog.md in full. For audits or thorough requests, also fetch the live page and note any signs not yet in the snapshot.

2. Pick the mode

• Draft: writing new text.
• Rewrite: revising existing text. Preserve the author's facts and voice, and the structure where sound; strip only the tells. Do not inject new claims.
• Audit: the user wants an assessment, not edits. Report which signs appear, where, with brief quotes. Mind the false-positive caveats in the catalog (§9): formal prose or perfect grammar alone proves nothing. Do not edit unless asked.

3. Write or edit

Apply the full catalog. The highest-value rules:

1. State facts plainly, with is/are/has/was. Not "serves as", "stands as", "boasts", "features", "marks", "represents", or "refers to" openings.
2. Cut inflated significance: "pivotal moment", "testament to", "underscores its importance", "reflects broader trends", "rich cultural heritage", "evolving landscape". Replace generic importance-claims with a specific, checkable fact. Specificity is the whole game.
3. Kill trailing participle analysis. No sentence-final "-ing" clauses that editorialize: "…, highlighting the significance of X", "…, ensuring Y".
4. Avoid AI-vocabulary clusters: delve, crucial, pivotal, tapestry, landscape (abstract), showcase, underscore, intricate, fostering, garner, meticulous, vibrant, robust, boasts, align with, testament. One may be fine; clusters are the tell. See the word list.
5. Break the templates. No "not just X, but Y" or "It's not X, it's Y" parallelisms; no rule-of-three triads everywhere; no "Despite these challenges…" conclusions; no "In summary / In conclusion / Overall" wrap-ups; no "It's important to note".
6. Attribute precisely or not at all. No "experts argue", "observers have noted", "industry reports suggest". Name the source or drop the claim. Never imply many sources when there is one.
7. Never use a dash as punctuation. No em dashes (—), and do not sneak the same construction back in with an en dash (–) or a spaced hyphen ( - ); that swaps the character while keeping the tell. Restructure instead: a comma, a colon, parentheses, or two sentences. Hyphens stay only inside compound words, and en dashes only in numeric ranges. (House rule; it is stricter than the Wikipedia page, which only flags overuse.)
8. Format like a human. Sentence case headings. No bullet lists where every item is a bold label plus a colon. Bold almost nothing. No emoji decoration, no needless small tables, no horizontal rules before headings, no skipped heading levels. Prose first; lists only when the content is genuinely enumerable.
9. No chat artifacts. No "I hope this helps", "Certainly!", "Would you like…", "Here is a…" framing; no knowledge-cutoff disclaimers; no speculation dressed as fact ("details are not widely documented, but likely…"); no unfilled placeholders like [Your Name] or 2025-XX-XX.
10. Use the target medium's markup, never Markdown headed for a non-Markdown destination. Never invent citations, links, DOIs, ISBNs, or page numbers. Verify every reference you emit, and strip tool residue (utm_source=, turn0search0, oaicite, [cite: 1]).
11. Vary rhythm, and repeat words when natural. Reusing the same noun beats cycling synonyms. Mix sentence lengths. Commit to definite statements ("was the first") and use plain verbs (wrote, used, moved, tried, died).

4. Self-review pass

Before delivering, reread the output against the checklist above. Scan it against the term lists in references/words-to-avoid.md (they are written for reading, not for grep -f). Two checks are mechanical; run them on the draft file:

# Dash punctuation and machine artifacts: any hit outside quoted material is a bug.
# (Unspaced en dashes in numeric ranges like 2024–2025 are allowed and not flagged.)
grep -nE '—| – | - |turn0|oaicite|oai_citation|contentReference|attributableIndex|attached_file|grok[_-]|(start|end)_span|ppl-ai-file-upload|【|\[cite:|\[web:[0-9]|:::writing|utm_source=(chatgpt\.com|openai|copilot\.com)|referrer=grok\.com|↩|XX-XX|INSERT_|PASTE_' draft.md

# High-signal AI vocabulary: hits are prompts to reread the sentence, not automatic errors
grep -inE 'delve|deep dive|tapestry|testament|pivotal|crucial|underscor|showcas|foster(ing|s)|garner|meticulous|vibrant|intrica|boasts|bolster|interplay|stands as|serves as|not just|not only' draft.md

Fix a hit by rewriting the sentence, never by thesaurus-swapping the flagged word. The underlying pattern (vagueness, puffery, formula) is the problem, not the word itself.

5. Report

• Draft/Rewrite: deliver the text. If rewriting, summarize which signs you removed in one or two sentences.
• Audit: list the signs found with locations and quotes, state an overall read, and note the base-rate caveat: these are statistical signs, not proof.

Fallback without network

Everything works offline from references/signs-catalog.md and references/words-to-avoid.md. Only the live-refresh step is skipped.

Maintaining the snapshot

When asked to update this skill from Wikipedia (or when the snapshot looks stale), reconcile semantically; the catalog is a reorganized distillation, so a textual diff against the wikitext is noise.

1. Check whether the page changed since the pinned revision:

   curl -fsSL 'https://en.wikipedia.org/w/api.php?action=query&titles=Wikipedia:Signs_of_AI_writing&prop=revisions&rvprop=ids|timestamp&format=json'
   

   If revid equals the revision pinned under "Source of truth" above, nothing to do.

2. Otherwise fetch and read the full current wikitext:

   curl -fsSL 'https://en.wikipedia.org/w/index.php?title=Wikipedia:Signs_of_AI_writing&action=raw' -o /tmp/signs.txt
   

3. Compare it section by section against references/signs-catalog.md. Fold in new signs, drop removed ones, and update references/words-to-avoid.md to match.

4. Update the pinned revision ID, timestamp, and sync date in both places: under "Source of truth" in this file and in the header of references/signs-catalog.md. The two pins must match.