The Style Guide

A portable Claude skill that holds multiple named writing / style guides in one place. Drop it into any AI session that needs to draft, ghostwrite, or edit text in a specific voice (or generate questions in a specific style, or strip a specific unwanted style) without hand-coaching the model every time.

Named for Michael Coté because his prose style is the shipped default, but the skill is designed to hold other styles too. Bring your own.

What's in here

Shipped styles:

• Coté - his prose voice across content types (blog / newsletter / personal LinkedIn / podcast prose / white paper / analyst brief / conference abstract / deadpan-list catalogue-of-the-absurd). Default when no other style is named.
• Tyler Cowen questions - question-generation style for interview prep in the mode of Conversations with Tyler: cite everything, jump domains, ask about tacit knowledge, overrated/underrated segments, personal biographical detail.
• LinkedIn - two peer files:
  • no-linkedin-talk - a negative / edit-pass style. Load when you're stripping LinkedIn clichés from a draft. Doesn't produce a voice on its own.
  • successful-posts - a positive / how-to style grounded in engagement research. Load when you want a LinkedIn post that actually performs. Ignores taste; pair with no-linkedin-talk if you want both.

Shared references (style-agnostic, loaded as-needed):

• formatting - dashes (never em-), quotes (never smart), italics (_underscore_, never *asterisk*), never **bold** for emphasis, the <figure> HTML pattern for uploaded images.
• no-signposting - the biggest AI-writing tic to strip on edit passes: "What's striking about X is...", "The thing that matters here is..." Worked before/after examples.
• ai-writing-tells - broader catalog derived from Wikipedia's Signs of AI writing project page. Has a 90-day freshness rule - if the file's Last fetched date is stale, prompt for a refresh.
• dont-talk-like-an-ai and dont-talk-like-an-ai-short - the anti-AI rules packaged to work with no skill at all. See the next section.

Two files you can use without the skill

references/dont-talk-like-an-ai.md and references/dont-talk-like-an-ai-short.md are written to be lifted out of this repo and used on their own. They're self-contained: no cross-references to other files here, no skill-loading machinery, no assumption that anything else got read first. Copy either one into a chat session, a system prompt, a CLAUDE.md, a .cursorrules, or the custom-instructions box of whatever assistant you use, and it works.

The long one runs about 200 lines and covers banned phrases grouped by failure mode, banned words, nine banned sentence structures, paragraph and document shape, punctuation, and a final-pass checklist. Use it when you're setting up a drafting or editing context and have room for it, or paste it into a single session when you want one piece cleaned up.

The short one is about 380 words, sized for a user-preferences box where you're competing with everything else you want the model to know about you. It keeps the highest-frequency bans and the structural rules that change how a reply reads, and drops the explanations. It opens with a line scoping the rules to every reply rather than only to documents you explicitly asked to have written, which is the difference between a model that stops saying "load-bearing" in chat and one that only stops when it thinks it's Writing.

Both are CC0 like the rest of the repo, so no attribution needed. If you do want the full skill, these two files are also loaded as references inside it, and styles/ai-detector.md plus scripts/cliche-check.py give you the same material with the sources and an automated scanner.

Layout

src/the-style-guide/
  SKILL.md                            # always-load: skill purpose, resolution order, XDG paths
  styles/
    cote/                             # multi-file style: general.md + register siblings
      general.md
      casual-voice.md
      professional-voice.md
      deadpan-list-essay.md
      social-media.md
    tyler-cowen-questions.md          # single-file style
    ai-detector.md                    # single-file style, voice-neutral edit pass
    linkedin/                         # two peer styles that share a subject
      no-linkedin-talk.md
      successful-posts.md
  references/
    formatting.md
    no-signposting.md
    ai-writing-tells.md
    dont-talk-like-an-ai.md           # standalone, paste anywhere
    dont-talk-like-an-ai-short.md     # standalone, sized for a prefs box
  scripts/
    cliche-check.py                   # regex scanner for the phrase lists

How to use it

Fastest path: clone the repo and run ./build.sh --install. That copies src/the-style-guide/ into ~/.claude/skills/the-style-guide/ so Claude Code picks it up. Set SKILL_INSTALL_DIR to install somewhere else.

If you'd rather do it by hand, drop src/the-style-guide/ into your Claude (or other AI) skills directory as the-style-guide/. Then either:

• Reference it in a session: "use the-style-guide skill for this draft."
• Or paste SKILL.md into the system prompt / context for one-off use.

Progressive disclosure: SKILL.md is always loaded. Individual style files load only when the user's request matches (case-insensitive substring match on filename stems, with hyphens / spaces / underscores treated as equivalent). References load only when a style points at them or the task warrants it. Only one style is loaded per task - asking for Tyler Cowen questions does not drag the Coté files into context.

If no style is named and the task is prose, the default is Coté. If a named style doesn't exist, the skill lists what's available and asks rather than silently falling back.

Customizing without forking

Follows the XDG config convention. Drop your own files into ~/.config/io.cote.ai.skill.the_style_guide/ (or set $THE_STYLE_GUIDE_CONFIG_DIR to point elsewhere) to layer on top of the shipped defaults:

• SKILL.local.md - standing rules applied to every task. Useful if you're using this skill in your own voice adjacent to Coté's, or you want to disable rules that don't apply to you. Contradictions win for the user.
• styles/<name>.md or styles/<name>/ - your own named styles. Same shapes as shipped. Local styles win on collision.
• references/<name>.md - static background docs.

See SKILL.md "Local customizations" for full resolution order.

Adding new styles

Two shapes:

• Single-file (styles/<name>.md) - one file covers the whole style. Good for compact styles. See tyler-cowen-questions.md for an example.
• Multi-file (styles/<name>/) - general.md is the always-load base when the style is picked; content-type siblings (blog.md, professional.md, etc.) load on top when the task matches. See styles/cote/ for an example.

A style file should cover: what the style is, sentence-level or question-level patterns, word choice (use freely / never use), structural defaults, anti-patterns, and a self-edit checklist.

Content notes

• Examples in the Coté style files include casual profanity ("shit", "fucked up") because that's the voice being encoded.
• All third-party sources (Wikipedia's Signs of AI writing, Onwuka's LinkedIn clichés post, Mollick's LinkedIn mute-list post, Usera/Cox/Walker's SAGE Open paper) are cited with URL and retrieval date in the source notes at the bottom of each file that uses them.

Who wrote this

Coté and Claude, working together across a lot of sessions. The rules come from Coté - his taste, his bans, the things that make him wince in a draft. Claude did most of the drafting, catalogued the sources, wrote the scanner, and got told when it was wrong. Commits carry a Co-Authored-By: Claude trailer where that's what happened.

There's an obvious joke here about an AI maintaining a list of ways not to sound like an AI. The list is better for it - a lot of these rules exist because Claude produced the tic first and Coté banned it after.

License

CC0 / public domain. See LICENSE. Copy, modify, redistribute. The point is portability - if this is useful to anyone else trying to encode a writing style for an AI to consume, help yourself.