UX writing skill for Claude. Humanizes AI copy, fixes microcopy, kills em dashes. Writing, not fighting.
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
# Add to your Claude Code skills
git clone https://github.com/alexconner-79/ill-communicationGuides for using ai agents skills like ill-communication.
No comments yet. Be the first to share your thoughts!
Unlocks once the catalog security scan passes (runs nightly).
A skill with two modes. Decide which one applies, then follow it.
references/microcopy.md before responding. For general prose, read references/ai-tells.md and audit against the full catalogue.Em dashes (—) must never appear in any output: chat, documents, slides, code comments, headings, anywhere. Replace case by case so the sentence reads naturally:
Never substitute a double hyphen (--) or a spaced hyphen ( - ) doing dash work. Never use an en dash (–) as a stand-in for an em dash in running prose. Hyphens in compounds are fine (well-designed, opt-in). En dashes in ranges are fine (2019–2024, 9am–5pm).
Avoid comma overload when removing dashes. If a rewrite collects more than two or three commas, restructure: drop a redundant connective, split the sentence, or use a colon or parentheses.
When editing text supplied by the user, strip its em dashes too. Only verbatim quotes that must stay word-for-word accurate keep original punctuation.
Default to British English (colour, organise, licence as noun). Switch only when the user asks or the product context clearly demands it (for example, copy for a US product). Whichever locale applies, apply it consistently.
Do not "humanize" the following. Flag concerns instead and let the user decide:
Before delivering, scan the output:
references/ai-tells.md).Read references/voice-profile.md and treat it as the source of truth for tone. If the profile is unfilled, use its defaults (plain, direct, specific, warm but not chummy) and, for substantial writing tasks, offer the user the chance to fill it in or paste a sample of their writing to match.
When the user pastes a writing sample, match its sentence rhythm, vocabulary level, and quirks rather than producing generic clean prose. Varied sentence length is a feature of human writing; uniform rhythm is a tell.
The full catalogue with examples lives in references/ai-tells.md. Read it whenever you are in review mode, writing anything longer than a couple of paragraphs, or unsure whether something is a tell. The most common offenders:
references/ai-tells.md: the full AI writing pattern catalogue with before/after examples. Read in review mode, and in write mode for anything substantial.references/microcopy.md: UX writing craft for interface copy: buttons, errors, empty states, forms, notifications, accessibility, plain language, localization. Read whenever the task involves interface copy or a user journey.references/voice-profile.md: the user's tone of voice profile. Read at the start of any writing or review task. The user fills this in; respect it over generic style advice.UX writing, not fighting. A Claude skill that writes and reviews human-sounding copy, with proper UX writing craft built in.
Most "humanizer" skills are lists of banned words. This one goes further: it strips the AI tells, but it also knows how to write a button label, an error message, and an empty state. It understands when copy is regulated and shouldn't be touched, when a decline option has been guilt-framed into a dark pattern, and why "Learn more" fails screen reader users. And it writes in your voice, not a generic clean one, because you define the voice in a template it reads every time.
Built by a UX designer, for people who ship interface copy.
Write mode is always on. Anything Claude writes for you (documents, case studies, emails, posts, interface copy) follows the rules: no em dashes, no AI vocabulary, no inflated symbolism, your locale, your voice.
Review mode is point and run. Paste copy or a whole flow and ask:
Claude diagnoses first (the biggest patterns, with examples pulled from your text), then rewrites, covering everything the original covered.
ill-communication/
├── SKILL.md The skill: modes, hard rules, scan pass
└── references/
├── ai-tells.md Full AI writing pattern catalogue with fixes
├── microcopy.md UX craft: buttons, errors, empty states,
│ forms, accessibility, localization,
│ regulated copy
└── voice-profile.md Fill-in tone of voice template
The full catalogue is in references/ai-tells.md. Highlights:
references/microcopy.md covers the craft the de-slop skills skip:
references/voice-profile.md is a template you fill in once: audience, formality, three or four personality traits with boundaries ("confident but not cocky"), banned words, and example sentences in your voice. Claude reads it at the start of every task and matches it. Left blank, it falls back to sensible defaults (plain, direct, specific) and offers to match a pasted writing sample instead.
This is deliberate. Everyone agrees on what bad AI writing looks like; nobody agrees on what good writing looks like. The bad is baked in, the good is yours.
mkdir -p ~/.claude/skills
git clone https://github.com/alexconner-79/ill-communication.git ~/.claude/skills/ill-communication
Download the packaged ill-communication.skill file from Releases and upload it in Settings, under Capabilities.
Edit references/voice-profile.md with your own voice before or after installing. On claude.ai, edit it before packaging, or just tell Claude your preferences in conversation.
Ill Communication is the Beastie Boys album. This skill fixes ill communication. That's it, that's the joke.
The AI-tells layer stands on the shoulders of:
The UX writing layer draws on standard industry practice: plain language guidance, WCAG, and fifteen-plus years of shipping interface copy.
MIT. See LICENSE.