Someone replied to an email last week saying it was one of the clearest I'd sent. I'd written maybe ten words of it. Claude wrote the rest. The thing that made it sound like me wasn't prompting skill or a special model. It was a rules file I set up once and haven't thought about since.
Here's what it is and how to build yours.
What a communication style file actually does
When Claude drafts something, it defaults to generic professional language. Clear, technically correct, nobody's voice. A communication-style.md file is a set of rules Claude reads before writing anything: your tone, your banned words, your sentence rhythm. Once it's in place, every draft starts from your voice instead of from scratch.
What to put in it
The file is plain English. No code. Four sections are worth having:
Tone. How you write in different contexts. "Internal messages: casual and direct. Client emails: professional but plain-spoken, never corporate." The more specific, the better.
What to avoid. This is the most important section. List the words and patterns you hate seeing in your own writing. Em dashes. Emojis. Phrases like "circle back", "touch base", "moving forward". The AI-tell lines that make it obvious a machine wrote it: "I hope this finds you well", "please don't hesitate to reach out", "as per our conversation". Kill them all by name.
Sentence rhythm. How your sentences should feel. Contractions? Short sentences or longer ones? How much variety? "Contractions are good. Vary sentence length. Avoid strings of three-word fragments back to back."
Default. What's your baseline? "When in doubt: shorter, sharper, more concrete. Strip filler. Trust the reader."
How to set it up (5 minutes)
- Open your Claude Code project folder. If you haven't set one up yet, the free template at github.com/alexsidhu1/execassistant-template gives you the full structure including this file.
- Create the
.claude/rules/folder inside your project if it doesn't exist. - Create a new file:
communication-style.md. - Fill it in. Copy the structure above and replace every section with your actual rules. Don't be vague: "professional" tells Claude nothing. "No em dashes, no emojis, never open with I hope this email finds you well" tells it exactly what to cut.
- Add one line near the top of your
CLAUDE.md:@.claude/rules/communication-style.md. This tells Claude to read your rules at the start of every session. If you're using the execassistant template, this line is already wired in.
That's it. From now on, every time you ask Claude to draft a reply, it reads your rules before it writes a word.
Using it to draft an email
The prompt is simple. Paste the email you're replying to and say:
Draft a reply. Keep it short. [Add any specific points you want to hit.]
Claude reads your style rules first, so it knows not to use em dashes, not to open with "I hope this finds you well," and to match your rhythm. Review the draft, tweak what needs tweaking, send.
You'll make fewer edits over time as the rules get sharper. When you see something wrong in a draft, add it to the file. "Never say 'as discussed'." "Don't use bullet points in client emails." Each addition compounds.
The honest ceiling
This setup is powerful for one person. Your voice, your rules, your project. The limitation: it lives in a folder on your computer. Everyone on your team has their own setup, or no setup at all. Company emails still sound different depending on who sent them.
The next level is a team-wide voice standard baked into a shared system, so every person drafting a client email starts from the same baseline. Same tone, same banned phrases, same rules enforced automatically. That's what we build into the AI operating systems at Whitehorse. If you want that for your business, an audit is where it starts.
Start here
Open a blank file and write your ten most-hated phrases. The words that make you wince when you see them in an email draft. Get those in the file today, and the next email Claude writes for you won't have a single one of them.
Built by Alex Sidhu, Whitehorse AI.