avoid-ai-writing

Audit & rewrite content to remove AI writing patterns. A practical skill for any AI agent. Supports detect-only and edit-in-place modes, plus voice profiles.

---

A portable writing skill for Claude Code, OpenClaw, Hermes, and any other agentskills.io-compatible agent. Audits and rewrites content to remove AI writing patterns ("AI-isms").

Three modes:

• Rewrite (default) — flags AI patterns and rewrites the text to fix them. A built-in second pass catches patterns that survived the first edit.
• Detect — flags AI patterns without rewriting. Shows which flags are real problems vs. judgment calls. Useful when patterns might be intentional, when auditing content you don't want altered, or when you just want a quick scan.
• Edit — edits a prose file in place (via the Edit tool) with minimal, targeted changes, preserving passages that are already human. Source code, configuration, and generated data are refused because prose rewrites can corrupt structured content. Returns an edits-made + verification report, not the full file.

Use --iterate N when you want the skill to repeat the audit and rewrite cycle until no patterns remain or the requested pass limit is reached. The limit is capped at 2: the initial rewrite plus one corrective pass. Rewrite mode already includes that corrective second pass; --iterate does not add passes on top of it. The skill reports how many passes it took.

An optional voice profile (casual / professional / technical / warm / blunt) sets how the prose should sound, independent of the audience context profile.

Quick demo

Input:

Certainly! Acme Analytics, a vibrant startup nestled in the heart of Boulder's thriving tech ecosystem, has secured $40M in Series B funding — marking a watershed moment for the observability landscape. The platform serves as a unified hub, featuring real-time dashboards, boasting sub-second queries, and presenting a seamless integration layer. Moreover, experts believe Acme is poised to disrupt the market. In conclusion, the future looks bright!

Output:

Acme Analytics, a Boulder-based startup, raised a $40M Series B. Its observability platform has real-time dashboards, runs queries in under a second, and includes an integration layer.

What it caught: the chatbot opener ("Certainly!"), promotional modifiers, inflated significance, roundabout verbs, vague attribution, and the generic conclusion. The rewrite keeps the funding, location, and three product capabilities. It removes the unsupported market prediction without inventing an investor or an integration mechanism.

Why a skill, not just a prompt

A one-shot "make this sound human" prompt catches the obvious stuff. This skill is different:

• Structured audit — returns identified issues with quoted text, the rewrite, a change summary, and a second-pass audit in four discrete sections. You see exactly what changed and why.
• Two-pass detection — the second pass re-reads the rewrite and catches patterns that survive the first edit: recycled transitions, lingering inflation, copula swaps that snuck through.
• 112-entry word replacement table across 3 tiers + 10 Tier 3 phrases — not vibes-based. Every flagged word has a specific, plainer alternative. "Leverage" → "use." "Commence" → "start." Tier 1 matches flag unless a listed exception applies, Tier 2 words flag when they cluster, Tier 3 words flag only at high density. Tier 1 itself splits into 1A frequency markers (delve, tapestry) and 1B clarity edits (in order to, utilize) — same fix, but only 1A is evidence about how a passage was produced, and 1B is weighted lower so a wordiness fix cannot push a document toward an AI classification. Tier 3 phrases (multi-word boilerplate like "the integration of," "decentralized compute") flag on per-phrase repetition or when 3+ distinct phrases stack in one piece — the LLM-self-varies-boilerplate shape.
• 74 pattern categories — representative examples below, each with before/after. Includes structural detection (hashtag stuffing, bare-NP bullet lists, hedge-stacked predictions), AI-tool fingerprints (placeholders, citation markup, UTM params), rhythm/uniformity checks, conversational-register tells, and writer-side tests. The full catalog lives in references/patterns.md; this count is enforced against it in CI.
• Detect mode — flag patterns without rewriting. See which flags are real problems vs. judgment calls. Useful when patterns might be intentional or you're auditing content you don't want altered.
• Works across platforms — a directory-based skill runs in Claude Code, Cowork (as a plugin), OpenClaw, Cursor (as a ported rule), and other directory-aware agents. See the install paths below.

Installation & Usage

Install the complete skill directory

Use the plugin install below or clone the repository into your agent's skills directory. Keep SKILL.md with references/patterns.md: the entry file loads the catalog before auditing. The bundled scripts/, detector/, and examples/ provide optional mechanical verification.

For a single-file rules field, use dist/avoid-ai-writing.md. It includes every rule and profile, with manual fallbacks for commands unavailable outside the bundle. Do not copy the slim entry file alone. Older installers that fetch only root SKILL.md omit its required reference; use a directory install instead.

Claude Code

Option 1: Clone into skills directory

git clone https://github.com/conorbronsdon/avoid-ai-writing ~/.claude/skills/avoid-ai-writing

Option 2: Copy a self-contained file

Download dist/avoid-ai-writing.md and place it in any directory that Claude Code can read. Reference it in your CLAUDE.md:

- Editing for AI patterns → read `path/to/avoid-ai-writing.md`

Option 3: Use as a slash command

Create a command file (e.g., ~/.claude/commands/clean-ai-writing.md):

---
description: Audit and rewrite content to remove AI writing patterns
---

$ARGUMENTS

Read and follow the instructions in ~/.claude/skills/avoid-ai-writing/SKILL.md

Then use /clean-ai-writing <your text> in Claude Code.

Claude Cowork — install as a plugin

Cowork loads skills only from installed plugins — it doesn't scan ~/.claude/skills/, so a bare clone (the Claude Code steps above) won't be discovered there. This repo doubles as a single-plugin marketplace, so install it as a plugin instead:

/plugin marketplace add conorbronsdon/avoid-ai-writing
/plugin install avoid-ai-writing@conorbronsdon-skills
/reload-plugins   # or restart the session, to activate the skill

In the Cowork desktop app, do the same from Customize → Plugins → Add marketplace from GitHub (conorbronsdon/avoid-ai-writing), then install avoid-ai-writing. The skill auto-triggers from phrases like "remove AI-isms." New releases arrive when the plugin's version is bumped — run /plugin marketplace update to pull them.

The same plugin install works in Claude Code if you'd rather have a versioned, updatable plugin than the file clone above.

Prefer not to install a plugin? Copy dist/avoid-ai-writing.md into a folder connected to your Cowork session and tell the agent to follow ./avoid-ai-writing.md — works as a one-off, no auto-trigger.

OpenClaw

Option 1: Install from ClawHub

clawhub install avoid-ai-writing

Option 2: Clone into skills directory

git clone https://github.com/conorbronsdon/avoid-ai-writing ~/.openclaw/skills/avoid-ai-writing

Cursor

Drop the ported rule into your project's .cursor/rules/:

mkdir -p .cursor/rules
curl -o .cursor/rules/avoid-ai-writing.mdc \
  https://raw.githubusercontent.com/conorbronsdon/avoid-ai-writing/main/cursor-rules/avoid-ai-writing.mdc

See cursor-rules/README.md for activation globs and trigger phrases. Functionally identical to the Claude Code skill — same tier vocabulary, same context profiles, same modes.

Hermes

Drop the skill into Hermes's skills directory — it then appears automatically as /avoid-ai-writing, no registration needed:

git clone https://github.com/conorbronsdon/avoid-ai-writing ~/.hermes/skills/writing/avoid-ai-writing

OpenAI Codex

Codex reads Agent Skills in the same SKILL.md format. Put it in .agents/skills/ at the repo root, or ~/.agents/skills/ to use it across all your projects:

git clone https://github.com/conorbronsdon/avoid-ai-writing .agents/skills/avoid-ai-writing

Native ChatGPT and Codex plugin package

Published in the OpenAI Plugins Directory; install it there for ChatGPT or Codex.

The plugin package keeps the canonical SKILL.md as its editorial authority and provides seven focused Skills:

• avoid-ai-writing — uses the canonical audit and rewrite instructions
• avoid-ai-writing-router — routes mixed and multi-stage requests
• ai-writing-detector — runs the bundled deterministic detector
• voice-preserving-rewriter — rewrites returned text while retaining voice
• file-edit-in-place — edits only explicitly named files
• preservation-verifier — checks meaning and constraints before and after a rewrite
• false-positive-reviewer — reviews detector findings that need context

From a cloned checkout, build and validate the package with:

python3 scripts/package-openai-plugin.py . /tmp/avoid-ai-writing.zip --json
python3 scripts/validate-openai-plugin.py . --json

Other agents

The generated dist/avoid-ai-writing.md (or the Cursor .mdc port) drops into most tools' rules/skills location:

Tool	Where to put it
Windsurf	.windsurf/rules/avoid-ai-writing.md
Cline	.clinerules/avoid-ai-writing.md
GitHub Copilot (VS Code)	paste into .github/copilot-instructions.md
Claude.ai Projects	paste dist/avoid-ai-writing.md into the project's custom instructions
ChatGPT Custom GPTs	paste dist/avoid-ai-writing.md into the GPT's Instructions field

Triggering the skill

Once installed, ask your assistant to clean up AI writing:

• "Remove AI-isms from this post"
• "Audit this draft for AI tells"
• "Make this sound less like AI"
• "Clean up AI writing in this paragraph"

In rewrite mode (default), the skill returns four sections:

1. Issues found — every AI-ism identified, with the text quoted
2. Rewritten version — clean version with all AI-isms removed
3. What changed — summary of the major edits
4. Second-pass audit — re-reads the rewrite and catches any surviving tells

In detect mode, the skill returns two sections:

1. Issues found — every AI-ism identified, grouped by severity (P0/P1/P2)
2. Assessment — which flags are clear problems vs. patterns that may be intentional or effective in context

Trigger detect mode with: "detect," "flag only," "audit only," "just flag," "scan," or similar.

Pattern reference

Representative examples from the catalog — not the exhaustive list (that's references/patterns.md). The skill's human-facing prose catalog and the detector engine use different counts on purpose: the engine implements 53 type categories because it splits the vocabulary tiers and adds stylometric/fingerprint signals (punctuation distribution, function-word entropy, bypass-trick detection) that work as math over a document rather than as a rule you'd look up. The two are mapped in detector/CATEGORIES.md; don't "fix" one count to match the other.

Content Patterns

#	Pattern	Before	After
1	Significance inflation	"marking a pivotal moment in the evolution of..."	State the supported event; do not invent a date or purpose
2	Notability name-dropping	"cited in NYT, BBC, and Wired"	Keep supported citations; remove claims that they prove importance
3	Superficial -ing analyses	"symbolizing... reflecting... showcasing..."	Replace with specific facts or cut
4	Promotional language	"nestled within the breathtaking region"	Keep the supplied location; remove promotional modifiers
5	Vague attributions	"Experts believe it plays a crucial role"	Request the source, or flag the claim for verification
6	Formulaic challenges	"Despite challenges... continues to thrive"	Name the challenge and the response
7	Novelty inflation	"He introduced a term I hadn't heard before"	Keep the personal observation if it matters; do not invent an explanation

Language Patterns

#	Pattern	Before	After
8	Word/phrase replacements (3 tiers)	"leverage... robust... seamless... utilize"	"use... reliable... smooth... use"
9	Copula avoidance	"serves as... features... boasts"	"is... has"
10	Synonym cycling	"developers... engineers... practitioners... builders"	"developers" (repeat the clear word)
11	Template phrases	"a [adj] step towards [adj] infrastructure"	Describe the specific outcome
12	Filler phrases	"In order to," "Due to the fact that"	"To," "Because"
13	False ranges	"from the Big Bang to dark matter"	List the actual topics
14	Parenthetical hedging	"tools (like X and Y)"	Name them directly or cut

Structure Patterns

#	Pattern	Before	After
15	Formatting	Em dashes (— and --), bold overuse, emoji headers, bullet-heavy	Commas/periods, prose paragraphs
16	Sentence structure	"It's not X, it's Y" + hollow intensifiers + hedging	Direct positive statements
17	Structural issues	Uniform paragraphs, formulaic openings, too-clean grammar	Varied length, lead with the point
18	Transition phrases	"Moreover," "Furthermore," "In today's [X]"	"and," "also," or restructure
19	Inline-header lists	"Speed: Speed improved by..."	Write the point directly
20	Title case headings	"Strategic Negotiations And Partnerships"	"Strategic negotiations and partnerships"
21	Numbered list inflation	"Here are 7 reasons why..."	Cut to the 2-3 that matter
22	False concession	"While X has limitations, it's still remarkable"	State the real tradeoff
23	Rhetorical question openers	"What if there were a better way to...?"	Lead with the claim

Communication Patterns

#	Pattern	Before	After
24	Chatbot artifacts	"I hope this helps! Let me know if..."	Remove entirely
25	"Let's" constructions	"Let's explore," "Let's break this down"	Just start with the point
26	Cutoff disclaimers	"While details are limited in available sources..."	Find sources or remove
27	Generic conclusions	"The future looks bright," "Only time will tell"	Specific closing thought or cut
28	Stock reaction framing	"What surprised me most," "I was fascinated to discover"	Keep specific, authentic reactions; cut empty stock framing or make the reaction concrete
29	Reasoning chain artifacts	"Let me think step by step," "Breaking this down"	State conclusion, then evidence
30	Sycophantic tone	"Great question!", "You're absolutely right!"	Remove entirely
31	Acknowledgment loops	"You're asking about," "To answer your question"	Just answer directly
32	Confidence calibration	"It's worth noting," "Interestingly," "Surprisingly"	Let the fact speak for itself

Meta Patterns

#	Pattern	Before	After
33	Excessive structure	5 headers in 200 words, "Overview:", "Key Points:"	Merge sections, use specific headers
34	Rhythm and uniformity	All sentences 15–25 words, all paragraphs same length	Mix short/long, fragments, questions
35	Over-polishing	Every irregularity sanded away, perfectly uniform prose	Keep natural disfluency, varied rhythm
36	Rewrite-vs-patch threshold	5+ vocabulary flags + 3+ pattern categories + uniform rhythm	Advise full rewrite, not patching

Structural Detection (v3.4)

Added in v3.4 to catch LLM output that sidesteps the vocabulary tables by substituting synonyms but still leans on structural shapes detectors can identify. Crypto/web3/AI-infra content is where these patterns concentrate most heavily, but the rules generalize to any social-length post.

#	Pattern	Before	After
37	Tier 3 phrases (multi-word boilerplate)	"the integration of," "decentralized compute," "community-driven," "long-term sustainability" stacked across a piece	Replace the repeated phrase with a specific claim, or vary genuinely. Flagged per-phrase at ≥2 hits, or as a cluster when ≥3 distinct phrases appear
38	Future-narrative closers	"may become one of the most important narratives of the next market cycle"	Pick the falsifiable version. "X may exceed Y by 2027" is a prediction; the template form is not
39	Hedge-stacked predictions	"could potentially create," "may eventually unlock"	Pick one. Each hedge cancels the next
40	"Real/actual" adjective inflation	"real on-chain tokenomics," "actual reward sustainability"	Drop the empty intensifier and add the specific claim. Carve-out: "real on-chain settlement, not bridged IOUs" is honest contrastive writing — the AI tell is the unsaid contrast
41	Hashtag stuffing	15-tag trailing block: #AI #Crypto #Web3 #Innovation #FutureTech…	2-3 specific tags max, or none. Empirical threshold: 6+ tags is near-universal in LLM social output, rare in thoughtful human posts
42	Bullet lists of bare noun phrases	* Stable mining efficiency / Reliable pool connectivity / Optimized RandomX performance / Low failed share rates / Effective hardware utilization / Consistent thermal stability	Convert to prose, or rewrite each item as a full claim with a verb and a number. Carve-out: genuine list content (changelogs, parameter docs, ingredient lists) where bare NPs are correct

AI-tool fingerprints & later additions (v3.5–3.8)

#	Pattern	Before	After
43	Unfilled placeholders	[Your Name], [INSERT SOURCE], 2025-XX-XX	Fill in with real content or delete — shipped placeholders are a near-definitive tell
44	Chatbot citation markup	citeturn0search0, oai_citation, contentReference[oaicite:0]	Strip the markup token entirely
45	AI-tool URL parameters	utm_source=chatgpt.com, utm_source=copilot.com	Strip the tracking parameter; keep the URL if the link matters
46	Speculative gap-filling	"maintains a low profile," "likely began his career"	Cut the guess, or replace with a sourced fact
47	Hyphenated modifier stacking	"a high-quality, well-architected, future-proof solution"	Cut to the modifier that matters; the individual hyphens may be correct
48	Infomercial engagement hooks	"The catch?", "The kicker?", "Here's the thing."	Delete the hook, state the thing
49	Vocabulary diversity (low TTR)	Narrow, repetitive word range across 200+ words	Broaden the what — name specific things, cite specific cases
50	Self-labeling significance	"That last move is the contrarian one," "This is the interesting part"	Cut the label; let the explanation carry the weight, or reposition the item so it stands out on its own
51	List-label periods	- **Intros.** Years of conferences and operator network. (also unbolded: - Intros. Years of...)	Use a colon, not a period, on a list label: - **Intros:** years of conferences and operator network.

Conversational-register patterns (v3.15)

Added after a real-world exchange in which a maintainer called out an assisted-sounding GitHub issue reply with "I prefer to talk human to human." Both are judgment calls rather than regex-detectable (an ordinary human paragraph and an AI-generated one can look structurally identical; the tell is the register and the redundant context, not a fixed shape) — see the detector/CATEGORIES.md §C note for why a first attempt at a wall-of-text detector was reverted.

#	Pattern	Before	After
52	Wall-of-text replies	A 4+ sentence, sub-150-word reply delivered as one unbroken paragraph with no line breaks — the shape LLMs default to in issue/PR comments, chat, and DMs	Break at thought boundaries. One idea per line-group, the way a person actually types a reply
53	Recap-flattery opener	"Thanks for all the legwork here — the migration script and the rollback plan you worked through are what made this possible."	Substance first. If thanks is warranted, one plain clause without the recap: "Thanks for the legwork — this looks right to me"

Share-post framing (v3.20)

#	Pattern	Before	After
54	Lingering-attention claims	"The line I keep coming back to:", "I can't stop thinking about this," "this has been rattling around in my head all week"	Open on the thing itself. Carve-out: keep the frame when a reason follows ("I keep coming back to exit-voice because it predicts who quits")

Narrated candor (v3.21)

#	Pattern	Before	After
55	Narrated candor	"Two caveats I would rather flag than let you discover later:", "I want to be upfront:"	State the caveats. Judgment-only: the same words carry real content in conflict-of-interest disclosure ("in the interest of full disclosure, I own shares in…"), which a regex cannot separate from the empty frame

Unnecessary hyphenation (v3.24)

#	Pattern	Before	After
56	Unnecessary hyphenation	"research-impact aggregator," "code-base," "in real-time," "works out-of-the-box"	"research impact aggregator," "codebase," "in real time," "works out of the box." Preserve legitimate modifiers such as "real-time analytics"

Two writer-side tests round out the catalog (judgment checks, not auto-detected): paragraph-reshuffle immunity (can you swap two body paragraphs without breaking the piece?) and the treadmill effect ("what's actually new in this paragraph?").

Full Example

Before (AI-generated):

Certainly! Here's a comprehensive overview of Acme's Series B.

Acme Analytics, a vibrant startup nestled in the heart of Boulder's thriving tech ecosystem, has secured $40M in Series B funding — marking a watershed moment for the company and the observability landscape at large. The round was led by Sequoia, with participation from Andreessen Horowitz, Y Combinator, and Index Ventures, underscoring the robust investor confidence in Acme's vision.

The platform serves as a unified hub for engineering teams, featuring real-time dashboards, boasting sub-second query performance, and presenting a seamless integration layer. It's not just another monitoring tool — it's a paradigm shift in how organizations navigate complex distributed systems. Experts believe Acme is poised to disrupt the $15B observability market. Studies show that companies utilizing comprehensive monitoring solutions can ascertain issues 40% faster.

• 🚀 Performance: The platform streamlines incident response, empowering engineers to resolve issues faster.
• 💡 Scale: Acme fosters collaboration across teams, from frontend developers to SRE practitioners to platform builders to infrastructure engineers.
• ✅ Adoption: Customer adoption continues to accelerate, reflecting broader industry trends.

Moreover, Acme's approach showcases the potential of modern observability — symbolizing a shift toward proactive monitoring, reflecting years of R&D investment, and highlighting the intricate interplay between data collection and actionable intelligence. Despite challenges common in the crowded monitoring space, Acme continues to thrive. In order to fully capitalize on this momentum, the company plans to expand its go-to-market endeavors.

In conclusion, the future looks bright for Acme Analytics. Only time will tell how this game-changer will reshape the observability paradigm. Feel free to reach out if you'd like more details!

After (cleaned up):

Acme Analytics, a Boulder-based startup, raised a $40M Series B led by Sequoia, with participation from Andreessen Horowitz, Y Combinator, and Index Ventures.

Its observability platform gives engineering teams real-time dashboards, queries that run in under a second, and an integration layer. The platform helps engineers resolve incidents faster and supports collaboration across frontend, SRE, platform, and infrastructure teams. Customer adoption is accelerating. Acme has invested in R&D for years and plans to expand its go-to-market work in a crowded monitoring market.

Claims needing sources: The input attributes a prediction about a $15B market to unnamed experts and a 40% improvement in issue identification to unspecified studies. Verify those sources before publishing the claims. The rewrite leaves them out; it does not replace them with invented benchmarks, customers, or plans. The product and adoption statements above also remain supplied claims, not independently verified facts.

What the skill caught: chatbot framing, promotional modifiers, inflated significance, roundabout verbs, repeated role synonyms, empty contrast, list decoration, vague attribution, filler, and generic closing language. The rewrite preserves the named investors and supplied capabilities, consolidates the team list, and flags the unsupported market and study claims for verification.

That's 35+ AI tells.

Run the detector

The skill ships a deterministic, zero-dependency detection engine in detector/ — the same engine the rules above describe, as runnable code. It works in Node (>=18) and the browser with no build step.

It's also the single source of the numeric score: the skill itself (and detect mode) report which patterns are present and how severe (P0/P1/P2), and the engine is what turns those into one computed 0–100 score. There's deliberately no second, prose-estimated score in SKILL.md — one scorer, not two.

npm install avoid-ai-writing-detector

The npm package includes the detector runtime and CLI entry points, but not the repository's scripts/ utilities. Clone the repository to run those scripts.

const AIDetector = require("avoid-ai-writing-detector");
const { score, label, issues } = AIDetector.analyzeText("Your text here…");
console.log(score, label, issues.length);

Score a file or piped text from the command line

The package also ships a zero-dependency CLI:

npx --package avoid-ai-writing-detector avoid-ai-writing draft.md
cat draft.md | npx --package avoid-ai-writing-detector avoid-ai-writing --context technical

After a global install (npm install -g avoid-ai-writing-detector) the command is available as avoid-ai-writing directly.

It prints the complete analyzeText() result as JSON and exits 0; usage and I/O errors go to stderr with exit code 2. Run avoid-ai-writing --help for the --context and --source-mode options.

Gate prose in GitHub Actions or pre-commit

The repository also ships a deterministic gate that fails on finding count per file, not the composite 0–100 score. That keeps CI policy independent of score recalibration work such as #70.

# .github/workflows/prose.yml
steps:
  - uses: actions/checkout@v7
  - id: gate
    uses: conorbronsdon/avoid-ai-writing@main
    with:
      glob: "**/*.md"
      threshold: "6"
      context: technical

For long-lived production workflows, pin uses: to a release tag or commit SHA that contains action.yml.

The Action exposes step outputs via $GITHUB_OUTPUT:

• pass: 'true' when all scanned files are within threshold; 'false' on a threshold failure or operational error.
• total-findings: total count of deterministic findings across scanned files; unset on an operational error (exit 2).
• failed-files: count of files exceeding the threshold; unset on an operational error (exit 2).

Downstream steps can consume these outputs:

  - name: Report gate summary
    if: always() && steps.gate.outputs.total-findings != ''
    run: |
      echo "Pass: ${{ steps.gate.outputs.pass }}"
      echo "Total findings: ${{ steps.gate.outputs.total-findings }}"
      echo "Failed files: ${{ steps.gate.outputs.failed-files }}"

The underlying avoid-ai-writing-gate CLI also accepts --json to emit structured JSON on stdout:

{
  "schemaVersion": 1,
  "threshold": 6,
  "context": "technical",
  "sourceMode": "rendered-markdown",
  "pass": false,
  "totalFindings": 9,
  "failedFiles": 1,
  "files": [
    { "path": "README.md", "findings": 2, "pass": true, "types": ["em-dash", "tier1"] },
    { "path": "docs/guide.md", "findings": 7, "pass": false, "types": ["hedge-stack", "tier1", "tier2"] }
  ]
}

Top-level fields report schemaVersion, threshold, context, sourceMode, pass (boolean), totalFindings, failedFiles, and files (preserving scan order). Each file item reports path, findings, pass, and sorted distinct detector types. When no files match the input or glob, files is empty with pass: true.

threshold is the maximum number of deterministic findings allowed in each file. The shipped default is 6, chosen from the current human-control corpus using the same technical + rendered-markdown settings as the Action. Across 376 human corpus documents, threshold 0 rejected 118/376 (31.4%); threshold 6 rejected 7/376 (1.9%). Six is also at or above the observed 95th-percentile finding count in every represented register (the technical-blog register has only one corpus document, so that slice remains under-sampled). Set threshold: "0" explicitly when a project intentionally wants a strict zero-findings policy. This is a writing-quality baseline, not an authorship classifier calibration.

The default context is technical, and Markdown is analyzed with rendered-markdown source masking.

Pre-commit users can install the repository hook:

repos:
  - repo: https://github.com/conorbronsdon/avoid-ai-writing
    rev: main
    hooks:
      - id: avoid-ai-writing

Pin rev to a release tag or commit SHA in shared repositories. The hook scans staged .md / .mdx files with the same 6-findings corpus-backed default. Use args to override the threshold, context, or source mode:

      - id: avoid-ai-writing
        args: ["--threshold", "0", "--context", "technical", "--source-mode", "plain", "--"]

Pre-commit replaces the hook's default args: ["--"] when you provide args. End an overriding list with "--" to protect filenames that begin with a dash. The entry keeps its defaults; later options in args take precedence.

The gate only detects. Preservation validation still requires an original and a rewritten file and remains a separate command:

node detector/validate.js before.md after.md

When working from a cloned checkout instead of the published npm package:

npm test          # run the detector's fixtures (no deps to install)

const AIDetector = require("./detector/patterns.js");
const { score, label, issues } = AIDetector.analyzeText("Your text here…");

When the input is a Markdown source file, pass { sourceMode: "rendered-markdown" } to exclude initial YAML frontmatter and HTML comments from the score while keeping issue offsets aligned with the original file. Plain-text behavior remains the default.

See detector/README.md for the full analyzeText API and detector/CATEGORIES.md for the rule ↔ category map that keeps SKILL.md and the engine in sync.

Use the detector over MCP

avoid-ai-writing-mcp wraps the published detector as a local stdio MCP server. It exposes two read-only tools: score_text for a compact result and audit_text for the full list of findings, suggestions, statistics, and highlighted sentence regions.

claude mcp add avoid-ai-writing -- npx -y avoid-ai-writing-mcp@0.1.0

The server calls no model and sends no text to a network service. It intentionally has no rewrite tool; rewriting stays in the skill, where the agent can apply the full editorial rules and preservation guardrails. See the MCP repository for configuration examples for other MCP hosts.

The engine also ships a preservation validator. detector/validate.js compares a rewrite against its original and fails when the edit touched something it shouldn't have: a code block, YAML frontmatter, a blockquote, a table cell, inline code, a URL, a file path, the heading structure, or when the rewrite ends with more flagged patterns than it started with.

node detector/validate.js before.md after.md   # exits 1 on a preservation error

Does it pass its own pass? PROOF.md scores this repo's documentation with this repo's detector and publishes the result, including two defects the scan found in our own work. npm run self-scan reproduces it, and CI fails when a document drifts past its budget.

House style is a different job

This skill removes AI-writing tells. Enforcing a published style guide is the different job: it doesn't do that, and it ships no style guides of its own. The optional --style input takes a house-style config you supply: a register list the model applies, and a mechanics object whose checkable rules scripts/check-style.js verifies deterministically (quote form and Latin abbreviations gate the exit code; heading case, em-dash rate, and number spelling are advisory). This script requires a cloned checkout. examples/ has the schema. You can skip the input entirely and put your guide in your agent's context alongside a voice profile, as instructions rather than as a checked rule set.

From a cloned checkout, after a rewrite, node scripts/normalize-quotes.js draft.md --reference original.md prints prose marks normalized to the original document's convention; add --write to save it. Explicit --quotes straight|curly overrides inference. It shares the checker's Markdown protection. See the usage and limits.

If you want Google, Microsoft, Red Hat, or Salesforce style checked in CI, Vale already covers that. Its package registry carries Vale-compatible implementations of those four, alongside ports of proselint, write-good, and alex. The four style-guide packages are MIT-licensed, though the guides they implement are not always (see the audit below); the linter ports vary, and proselint's is BSD-3-Clause. The two tools do different jobs and compose: Vale gates a document against a rule set, applying fixes one alert at a time, while this skill rewrites whole passages as you draft.

Paywalled guides (Chicago, APA, MLA, AP) have no machine-readable implementation here or in Vale, and won't get one here. Nothing in this repo could verify that a rewrite is Chicago-compliant, so claiming it would fail the same bar PROOF.md holds every other number to. Passing one of their names to --style bundles nothing; it falls back to the model's own knowledge, and SKILL.md instructs it to say so and to claim no compliance. That is an instruction rather than a checked rule, which is the point: there is nothing here to check it against. The license audit behind that line is public.

More from me

I run agents against my own email, money, and publishing, so I build the guardrails first.

Follow me on GitHub for new tools and practical examples from maintaining them.

• What I'm building — public projects across agent skills, MCP servers, creator tools, and Mojo.
• Chain of Thought — conversations with the people shipping AI systems.
• repo-audit — checks whether repository claims are enforced, advisory, or guidance.

Credits

Pattern research informed by:

• Pangram Labs AI detection research — structural regularity insights, vocabulary flags from a decoder-only classifier trained on 28M human documents
• Wikipedia's Signs of AI-generated text documentation — the canonical reference for AI writing tells, maintained by Wikipedia editors
• blader/humanizer Claude Code skill
• brandonwise/humanizer — tiered vocabulary system, statistical analysis research (burstiness, sentence length variation, trigram repetition), and rewrite philosophy
• OpenClaw humanizer skill ecosystem — community patterns and vocabulary research

Pull requests get an automated first-pass review from Qodo Merge, free through Qodo's open source program. Thanks to the Qodo team for supporting OSS maintainers.

Authored by Conor Bronsdon · LinkedIn · Chain of Thought podcast

Community / Multilingual

Things the community has built around this skill:

• avoid-ai-writing-multilingual by Jürgen Kraus — German (SKILL-DE.md), French (SKILL-FR.md), Italian (SKILL-IT.md), Romanian (SKILL-RO.md), and Swedish (SKILL-SV.md) adaptations, grounded in native-language research rather than translated from English.

Built something on top of this skill? Open an issue — happy to link it here.

---

Disclaimer

This is an independent personal project, not affiliated with, sponsored by, or endorsed by any company. All views expressed are my own.

License

MIT