What a model needs from you before it can write as you

It's Tuesday. Someone needs a welcome email by Thursday and you are the closest thing this company has to a marketer. You open Claude and type the honest version of what you want: write a welcome email for our product, friendly and professional, not too salesy. Fifteen seconds later you have a competent email. Greeting, three benefits, a button, a sign-off. It is fine. It is also completely anonymous — you could paste your competitor's logo on it and nothing would look wrong.

That's not the model failing. That's the model doing precisely what you asked with the only information you gave it. "Friendly" is a word that describes millions of emails, so it points at the centre of all of them. Ask for the centre and you get the centre. The output is average by construction, and average is the one thing your welcome email cannot afford to be, because a new signup's entire read of your company right now is this email and a logo.

So the fix isn't a better adjective. Nobody has ever been saved by upgrading "friendly" to "warm but authoritative". The fix is replacing description with evidence — showing the model prose that is already yours, plus the specific moves you would never make. Everything below is how to assemble that in about forty minutes, once, for a document you'll reuse for a year.

Open whatever you use for internal chat and go looking for five to ten sentences you wrote, sent to a real human and would happily send again. Not paragraphs. Sentences. The best hunting grounds, in order of yield: the sales email that got a reply within an hour, the message where a founder explained the product to a confused customer at 11pm, the release note somebody screenshotted and said "this is good" about.

A sentence carries a dozen decisions that no adjective encodes. How long you let a clause run. Whether you use contractions. Whether you address the reader as "you" or talk about "customers" in the third person. Where you put the verb. Whether you capitalise your product features. Whether you ever use an exclamation mark and, if so, how many per year. Hand a model ten of those and it has a distribution to imitate that is narrower than the internet by roughly the entire internet.

Include one more thing that most people skip: a sentence you nearly sent and killed, with the reason. "We're thrilled to announce our newest feature — killed, because we are not thrilled and nobody believes we are." A rejected example with a stated reason teaches more than three accepted ones, because it draws the boundary rather than the centre. If you have two of those, use two.

Here is the asymmetry that decides how useful your brief will be. A model cannot check its own draft against "warm" — there is no test for warmth it can run. It can check its draft against "never open with a rhetorical question" in about a second, because that's a string search with a rule attached. Negative constraints are enforceable. Positive adjectives are aspirational.

So the largest block in your brief is the list of things you never do. Fifteen to twenty entries, no more, because a longer document is one nobody actually pastes. The file this site is written against bans the verb 'unlock' in front of anything abstract, the three-item rhythm AI prose falls into when it has nothing to say and the whole family of praise-words that sound like a specification and carry no information. Yours will be different, and the differences are the point.

Attach a reason to every ban. This is the part that separates a brief that works from a list that gets obeyed literally and evaded in spirit. "No 'effortless'" gets you a draft that says "frictionless" instead. "No 'effortless' — praise-words that don't describe anything specific; name the specific quality instead" gets you a draft that says what actually happens. The reason generalises to the hundred cases you didn't think to list. Orbit's own voice file is built this way: every banned pattern carries a why and an instead, and the instead is doing most of the work.

The Orbit Anti-Slop Editor skill carries the general-purpose version of this list — the patterns that mark prose as machine-drafted regardless of industry. Take it as your starting set and then add the ones specific to you, which are usually the phrases your own team has worn out without noticing.

Voice isn't a fixed property of a company. It's a relationship, and the same brand writes very differently to a trial user on day two than to a customer who cancelled in March. If your brief doesn't say who is reading, the model will invent a reader. The reader it invents is a generic prospect who knows nothing — which is why AI drafts constantly re-explain your product to people who have been paying for it for a year.

Write a short block per lifecycle stage — the phases a customer moves through, from signup to activation to steady use to lapse. For each one: who this person is, what they did in the last day, what they already know and what they're privately worried about. Three or four lines each. "Day 2 trial: has connected one integration, has not invited a colleague, is wondering whether this is going to take a week to set up." That single line changes the draft more than any tone instruction, because it tells the model what the email is for.

The strongest entries in this block read like audience notes but function as voice rules. "Never explain what the product does to someone who has been paying for a year" is an audience fact and a register instruction at the same time. So is "a churned customer does not want to hear that we've been busy shipping." Collect those as you notice them; they arrive one at a time, usually while you're wincing at a draft.

Create a file. brand-brief.md, in whatever folder you'll actually find again — the repo, the shared drive, a note you can paste from on a phone. Six blocks:

Then use it as the first thing in the session, before the request: paste the brief, then ask for the email. If you work in a tool that reads files, keep it on disk and point at it — the version that lives in a file gets updated, the version that lives in a chat history gets retyped slightly differently every time and slowly drifts into something nobody agreed to.

The last block is the one that keeps this honest, because a brief with no check is a document you wrote to feel organised. Run the finished draft through the slop detector, which scores prose against the machine-drafting patterns rather than against your taste. Orbit's own voice file sets 85 as the floor for anything that ships on this site; 70 to 84 is fixable with one editing pass, below 70 means start again. Pick your own floor if you like, but pick a number, because "does this feel like us?" asked at 6pm on a Thursday always resolves to yes.

One warning about what the brief can and can't do. It transfers a voice that exists. It does not create one. If you sit down to fill block two and cannot find five sentences you'd defend, the brief is not what you are missing. You have a writing problem wearing an AI costume, and the fix is to write five sentences to one real customer, badly and today, then start the brief from those. The brand voice guide covers how to decide what those sentences should sound like in the first place.

Forty minutes, six blocks, one file. Do it before the next email rather than after, because the draft you're about to write is going to set the tone for every draft that copies it.