Make AI writing sound like your company
AI writing style guide is an Agent Skill: a written method your AI applies on demand, in Zealos or in any client that reads SKILL.md.
AI writing style guide is an Agent Skill: a written method your AI applies on demand, in Zealos or in any client that reads SKILL.md.
A writing-voice skill is the highest-leverage thing you can give an AI, because it changes every sentence the product produces and it needs no tools, no integrations and no data.
| name | our-writing-voice |
|---|---|
| description | The house style for anything we write: emails, proposals, posts, docs. Use whenever drafting or editing prose that a customer, prospect or teammate will read. |
| allowed-tools | none — this skill adds judgement, not reach |
| compatibility | Zealos, Claude Code, Cursor, Codex and any client that reads SKILL.md |
| license | Free to use and edit |
--- name: our-writing-voice description: The house style for anything we write: emails, proposals, posts, docs. Use whenever drafting or editing prose that a customer, prospect or teammate will read. --- # Our writing voice Default to how a competent colleague writes when they respect the reader's time. ## Rules 1. **Lead with the point.** The first sentence says the thing. Context comes after, and only if the reader needs it to act. 2. **Short sentences beat clever ones.** If a sentence needs a second read, split it. 3. **Concrete over abstract.** "Cut onboarding from 9 days to 2" beats "dramatically accelerates time-to-value". 4. **Name the actor.** "We missed the deadline", not "the deadline was missed". 5. **One idea per paragraph.** Three sentences is usually plenty. 6. **No throat-clearing.** Cut "I hope this finds you well", "I wanted to reach out", "As you may know", "Per my last email". 7. **Ask for one thing.** If a message has two asks, it gets neither. ## Banned Leverage (as a verb) · utilize · synergy · seamless · robust · best-in-class · cutting-edge · game-changer · revolutionize · unlock · empower · elevate · supercharge · "in today's fast-paced world" · "it's not just X, it's Y". ## Calibration Read the draft aloud. If you would not say it to the person across a table, rewrite it.
Everything the product writes sounds like you wrote it, which matters most in the year when every email still is from you.
Brand voice stops being a deck nobody opens and becomes a document that is applied automatically to every draft, every time.
Drafts arrive needing a tweak rather than a rewrite, because the model already knows you do not open emails with I hope this finds you well.
Two things, applied in order: a voice skill that says how you sound, and a de-slop pass that removes the statistical tells — rule-of-three lists, hedging stacks, participial tails and unearned enthusiasm.
None. It is pure method, which makes it the clearest illustration of why skills and agents are separate concepts.
Yes. Keep a workspace voice for anything customer-facing and a personal one for your own drafts; personal skills are visible only to their author.
AI email triage is worth having only when it sorts by consequence rather than by recency: what blocks a deal, what someone is waiting on you for, and what can be read at the end of the week.
/email-rulesA follow-up gets a reply when it opens with the specific commitment made on the call, stays under 120 words, and asks for exactly one thing the recipient can answer in a sentence.
/email-rulesAI writing gives itself away structurally before it does lexically: paragraphs of identical length, lists that always contain three items, and a closing summary that restates a body the reader just finished.
/no-slopA proposal is read by people who were on none of the calls, so it has to open with what the buyer told you, in their words — get that visibly right and it buys the read for everything after it.
/proposal-structureAfter-call work is the twenty minutes nobody books: the CRM update, the recap, the follow-up, the task for the thing you promised — and it is the first work to disappear on a busy day.
/call-recap-formatA cold sequence works when each touch carries new information: the second email should not exist unless it says something the first one did not, and most sequences fail this test by touch three.
/email-rules