A Claude Code skill that audits and rewrites content to strip out AI writing patterns ("AI-isms") — the em dashes, hollow intensifiers, vocabulary tells, template phrases, hashtag stuffing, and metronomic rhythm that make text sound machine-generated.
Ships in the onex plugin — invoked as /onex:rewrite.
What it does
Paste a paragraph, a LinkedIn post, a blog draft, an investor email, or a piece of docs and trigger /onex:rewrite (or just say "remove the AI-isms," "make this sound less like AI," "humanize this"). Claude will:
- Audit it — flag every AI-ism with the offending text quoted, grouped into three severity tiers (P0 credibility killers, P1 obvious AI smell, P2 stylistic polish).
- Rewrite it — return a clean version that preserves the original structure, intent, and technical detail. Only changes what the guidelines require.
- Diff it — summarise the meaningful edits so you can see what moved and why.
- Second-pass it — re-read its own rewrite for anything that survived the first pass and fix it inline.
Two modes:
rewrite(default) — full audit + rewrite + diff + second pass.detect— flag only, no rewriting. Use when the patterns might be intentional, you don't want the text altered (someone else's writing, published content, reference material), or you just want a quick scan.
Seven context profiles adjust how strict the rules are: linkedin, blog, technical-blog, investor-email, docs, casual, landing-page. Auto-detected from cues (short + hashtags → linkedin, code blocks → technical-blog, salutation + fundraising language → investor-email, etc.), and you can override.
What it catches
A non-exhaustive sample of the rule categories — the full SKILL.md has the complete list with replacement tables:
- Vocabulary tells — three-tier word list (Tier 1 always flag: delve, leverage, robust, comprehensive, seamless, game-changer; Tier 2 flag in clusters: harness, foster, streamline, empower; Tier 3 flag by density: significant, innovative, compelling).
- Formatting — em dashes, bold overuse, emoji in headers, excessive bullet lists, title-case headings.
- Structural tells — uniform sentence and paragraph length (the #1 detection signal), formulaic openings, compulsive rule of three, copula avoidance (serves as, features, boasts), synonym cycling, excessive headers.
- Headline tells: counting frames and spelled-out numbers as hooks (One Job, One Fixed Fee, Three Jobs Teams Hand Off First), journey templates (From X to Y, Start small, see it work, then…), time frames (Live by Week 6, Week by Week), fee mechanics (Pay Once and Keep the AI Workflow), and Title Case. Rewrites follow how Apple, Stripe, and YC startup landing pages write headlines. The headline carries the payoff in the buyer's words, and the subhead carries the proof in digits. Section headings answer instead of label, each page keeps one voice, and a regex lint covers headlines, eyebrows, buttons and step titles.
- Significance inflation — watershed moment, pivotal moment, marking a turning point in the evolution of.
- Generic future-narrative closers — may become one of the most important narratives of the next market cycle.
- Hedge-stacked predictions — could potentially, may eventually, might ultimately.
- Real/actual adjective inflation — real on-chain tokenomics, genuine utility, true product-market fit (with a carve-out for named contrasts).
- Hashtag stuffing — 6+ hashtags on a short post; mix of project tag + broad category tags.
- Bullet lists of bare noun phrases — 5+ short adj+noun items with no verbs.
- Chatbot artifacts — I hope this helps!, Great question!, Let's dive in!, In this article, we will explore….
- "Let's" constructions — Let's explore, Let's take a look, Let's break this down.
- Vague attributions — Experts believe, Studies show, Research suggests (without naming the expert or study).
- Filler phrases — It is important to note that, In terms of, The reality is that.
- Chatbot citation markup leaks —
citeturn0search0,contentReference[oaicite:0],oai_citation(fingerprints from chat UIs that survive copy-paste). - AI-tool URL parameters —
?utm_source=chatgpt.com,?utm_source=claude.ai,?referrer=grok.com. - Unfilled placeholders —
[Your Name],[INSERT SOURCE URL],2025-XX-XX. - Reasoning chain artifacts — Let me think step by step, Breaking this down, First, let's consider.
- Confidence calibration stacking — It's worth noting that, Interestingly, Surprisingly, Notably, Undoubtedly.
- Emotional flatline — What surprised me most, What struck me was, The most interesting part.
- Novelty inflation — He introduced a term, the failure mode nobody's naming, what nobody tells you about.
How it's different from a generic "improve this writing" pass
- Pattern-specific, not vibes-based. Every flag points to a named category with a documented fix, not "this feels off."
- Severity-tiered. P0 / P1 / P2 lets you triage. Cutoff disclaimers and chatbot artifacts are credibility killers; uniform paragraph length is polish.
- Context-aware. A "thriving ecosystem" survives a casual Slack message and dies in an investor email. The same rule, calibrated by profile.
- Detect-only when you want it. You're not forced into a rewrite when all you wanted was a scan.
- Honest about false positives. The skill is explicit that these are signals, not proof — independent audits find 60%+ false-positive rates on non-native English writers. It's a quality tool, not a verdict tool.
Install
This skill ships in the onex Claude Code plugin:
/plugin marketplace add onextech/skills
/plugin install onex@skillsThen trigger it by saying /onex:rewrite, rewrite this, remove the AI-isms, humanize this, or any of the phrases listed in SKILL.md's description.
Credits
Adapted from the avoid-ai-writing skill by Conor Bronsdon (MIT, v3.4.0, agentskills.io spec 1.0). Vocabulary tiers draw on research from brandonwise/humanizer; chatbot citation-leak rules (P34/P35) are from Aboudjem/humanizer-skill. Repackaged for the onex plugin so it can be invoked as /onex:rewrite.
License
MIT