Style guide · Version 3

The LLM style and copy guide

A free set of writing rules you give to ChatGPT, Claude, Gemini, or any other AI chatbot. Upload it once, and the drafts that come back use plain words, keep your facts, and drop the habits that make people stop reading.

Pick an edition See it on GitHub

What it's for

Ask an AI tool for a post or an email and the first draft usually shows the same habits: em dashes everywhere, "delve" and "seamless," a cheerful intro, three tidy bullets, a recap of what you just read, and sometimes a number nobody measured. Readers notice, even when they can't say what's off.

The guide tells the model which words and sentence patterns to avoid, what to write instead, and how to run a six-check final pass before it hands the draft back. It's just as strict about what to keep: the facts, the limits, the exact names and button labels, and the next step your reader needs.

Before

"AI isn't just a tool. It's your unfair advantage. The people who embrace it today will be the ones leading tomorrow. Let that sink in."

After the style pass

"Before adding AI to your week, pick the one task you want to change. Drafting a reply from meeting notes is a specific task. 'Be more productive' is too broad to test. Try it for a week and notice how much editing the drafts still need."

It's an editing standard. It isn't an AI detector, and its word lists aren't proof that someone used AI. Plenty of people wrote "delve" long before chatbots did.

Who can use it

Anyone who uses an AI tool to draft something they'll send under their own name or their organization's name. You don't need to be technical. If you can upload a file to a chat, you can use it. It was written with these readers in mind:

Pick your edition

Every edition is complete on its own. Choose the one closest to what you write most and upload that single file.

Version 3

Core

  • Emails and messages
  • Social posts and newsletters
  • Bios, web pages, and instructions

The full rule set for everyday writing, with general writing modes and examples.

Version 3

Business

  • LinkedIn and professional accounts
  • Website, service copy, and case studies
  • Proposals, client email, and listings

Startup language to rewrite, business claims that need proof, and calls to action that serve the reader.

Version 3

Tech

  • Explainers and news roundups
  • Tool reviews and release notes
  • Setup docs and security notices

Rules for news, AI features, and security copy, plus a tech hype table.

Version 3

Education

  • Lesson materials and directions
  • Student feedback
  • Family newsletters and staff messages

School terms, three school audiences, and rules against inventing student observations or progress.

How to use it

  1. Fill in Make it yours

    The table near the top of each edition sets whose voice it is, who reads it, three words for how you want to sound, and the words you never want to see. Anything you leave blank falls back to the defaults.

  2. Upload it to a project

    Add the file to a Claude Project, a ChatGPT Project or custom GPT, or a Gemini Gem. If your tool doesn't have projects, paste it into custom instructions or at the start of a chat.

  3. Start each task with the instruction

    Every edition ends with a starter instruction. Fill in the brackets and paste it in with your notes. Here's the core version:

Use the LLM style and copy guide, version 3 core edition, for this task. Write for [audience] in [format or platform] as [my name, or the organization's name]. The purpose is [what the reader should understand or do]. Use my notes as the source of facts and keep the substance, constraints, and necessary detail. Use the voice and terms in Make it yours. Replace generic praise, hype, engagement bait, and stock LLM patterns with concrete wording. Don't invent experience, testing, quotations, credentials, features, or results. Verify facts that change over time when they matter, and label anything still missing. Include a call to action only when it serves the purpose. Keep editorial flags separate from the finished copy. Run the guide's final pass before you return the draft.
  1. Read the draft yourself

    The final pass catches patterns. It can't check your facts, your dates, or whether the story in the draft really happened. That part stays with you.

What's in every edition

Each file also starts with skill frontmatter, so you can drop it into a folder, rename it SKILL.md, and use it as a Claude skill.

Version history

Version 3, October 2026
Four editions (core, business, tech, education), the Make it yours table, format notes for each place you publish, stricter evidence rules, expanded phrase tables, and new worked examples.
Version 2, July 2026
Sentence-pattern bans, rules against reflexive praise and "I hope this helps," evidence rules, a longer vocabulary list, and the six-check final pass.

The full list of changes is in the changelog.

Free and open source

The guide is MIT licensed. Copy it, adapt it, and put your name on your version. If you find a pattern it misses, or a rule that gets in the way of good writing, open an issue on GitHub or send it in here with an example and a rewrite. A preference that's only yours belongs in your own Make it yours table.

Credit and references

Created by Amanda Grutza for how to talk to robots. For another open-source take on the same problem, see Peter Yang's no-ai-slop skill, which removes AI writing patterns from a draft.