Setting up your agent

Your voice

Real writing samples beat adjectives. How the agent learns to sound like you.

Voice is the hardest thing to describe and the easiest thing to show. "Direct, warm, a little dry" describes half of LinkedIn. Three paragraphs you actually wrote describe only you. Everything in this part of the setup is built around that fact.

Samples first

The best voice input you can give the agent is screenshots of posts you have already written, uploaded at the head-start step. The agent reads the text out of them once, keeps the text as a note called "My past LinkedIn posts", and discards the images. From real posts it picks up the things nobody thinks to write down: sentence length, whether you open with a question or a statement, how you use line breaks, whether you ever swear, how you close. That is why a LinkedIn profile PDF, which contains no posts, gives the agent your history but not your voice — and why the voice question is never pre-filled from one.

Don't have posts? Anything you wrote in your own words works: an email you are proud of, a talk transcript, a newsletter, a long comment. Paste it into the voice answer or add it as a Text source in your material.

Then describe it — briefly

The voice question is optional and works best as a short list of concrete habits rather than a personality profile. Habits the agent can act on: "short paragraphs, one idea each", "I never use exclamation marks", "I open with the conclusion", "I say 'the IRS' never 'the Service'", "no bullet points ever".

Rich enough to write from

"Plain and calm. Short sentences. I explain the thing, then say what I'd do. First person, 'I' not 'we'. No rhetorical questions, no 'Here's the thing', no emojis. I'm allowed one dry joke per post, max. I almost always end on the practical next step, not a moral."

Every line is a rule the agent can follow and you can check. The "one dry joke" line is worth more than the word "witty" would have been.

Too thin to write from

"Professional, knowledgeable, approachable, authentic."

Four adjectives that rule out nothing. The agent falls back on the generic professional register — which is exactly the voice you are trying to avoid.

What you would never sound like

The next question — what makes you cringe — is often more useful than the voice itself, because it draws the boundary. Name the writers, phrases and moves you would never use: "hustle content", "I'm humbled to announce", posts that open with a one-word hook, fake vulnerability, emoji bullet points, anything that sounds like a TV ad for tax relief. The agent treats these as hard rules, and a thumbs-down with "this is exactly what I said I hate" is the fastest correction there is.

Off-limits topics

Anything the agent should never touch: politics, specific clients or cases even anonymised, your competitors by name, your family, a former employer, your fees. Also anything you cannot say for professional reasons — a lawyer might rule out anything that could read as advice on a specific matter. This list is honoured absolutely; it is not a preference the agent weighs.

What the agent does with all this

Every draft is written against your voice notes and your samples, and checked against your never-list before you see it.
Recent posts you rated are shown to the agent as examples — good ones to keep sounding like, bad ones with your note on why. See Feedback that changes future posts.
Your voice is yours. Nothing you provide is used to train shared models or to write for anyone else. See What we store.

If you only have five minutes: skip the voice description entirely and upload screenshots of your five best posts. Then grade the first drafts. That combination beats any description you could write.