How-to
How to Write a Brand Profile for an Agent
Stop re-prompting your agent on voice: write the brand profile once.
A brand profile is the eight-field document your agent reads before every run. Here is how to write each field so your content stays on-voice without re-prompting.
The draft is clean, and it sounds like nobody
The draft comes back clean. Grammar is fine, the structure is sound, the facts check out. And it sounds like nobody. It has the cadence of every other post in the feed: the throat-clearing intro, the breezy confidence about a topic you cover with more care than that, the words you would never actually use. You did not write a bad prompt. You wrote a fine prompt, and you will write another fine prompt tomorrow, and the day after, because the voice keeps sliding back to the mean every time the context resets. The problem is not the request. The problem is that the agent has no standing record of who you are.
The standing brief
Here is the fix, and it is boring on purpose: write your editorial standard down once, in a structured profile the agent reads before every run. Call it the standing brief. A brand profile is not a vibe or a tone slider. It is a field set, eight categories, that encode the decisions an editor would otherwise make in their head: how you sound, what words you use and refuse, whose attention you are not chasing, what each channel allows, and where a reader goes next. Written once, persisted, and loaded into every run, it does the job that re-prompting only pretends to do.
Why 'just prompt better' keeps failing
The common advice is to prompt better. Add more adjectives, paste in three sample posts, tell it to write like a seasoned analyst and not a marketer. That works for one draft. It fails on the second because nothing carried over. Prompting is the wrong layer for voice the same way retyping your address on every form is the wrong layer for an address book. Voice is not a per-request instruction; it is a property of the publisher that should outlive any single task. Treating it as something you re-explain each session guarantees drift, and drift is what readers feel when a post is technically yours but tonally a stranger's.
The eight fields, and the question each one answers
The profile has eight fields, and each answers a question an editor asks reflexively. Voice: how you sound in three or four sentences, the register and the rhythm. Lexicon: the words and phrases you actually use, your terms of art, the way you name things in your niche. Framing: the lens you argue from, your recurring thesis, the angle that is yours. Anti-ICP: who you are not writing for, so the agent stops hedging toward a general audience. Channel constraints: what each surface allows, length, formatting, what a LinkedIn post can do that a newsletter cannot. Proof assets: the sources, data, and credentials you cite from, so claims are grounded in your evidence and not invented. Banned words: the terms that instantly read as off-brand, your personal salt list. Conversion path: where a reader should go next, the standing call to action that fits each piece.
Watch the fields fill in
Take a solo writer covering climate policy for municipal planners. Watch the fields fill in. Voice: plain, precise, a little dry; short sentences; we explain mechanisms, not outcomes, and we never sound alarmed. Lexicon: we say 'jurisdictions' not 'communities,' 'capital plan' not 'budget,' 'adaptation' not 'going green.' Framing: every story answers what this means for a planner with a fixed capital plan and a council to convince. Anti-ICP: not activists, not investors, not a national-policy audience; a planner in a mid-size city. Channel constraints: the newsletter runs 600 to 900 words with one chart; LinkedIn is a single tight observation under 200 words with no link in the body. Proof assets: we cite primary sources, the ordinance text, the agency dataset, the council minutes, never a secondary summary. Banned words: 'unprecedented,' 'game-changer,' 'leverage,' 'in today's world.' Conversion path: each piece ends pointing to the relevant briefing in the archive. Now ask the agent for a LinkedIn post on a new federal rule. It already knows the length, the lexicon, the planner it is talking to, and the words to avoid. You did not say any of that this time.
What changes when the profile holds
When the profile holds, three things change. Drafts stop drifting, because the standard does not reset between runs. Repurposing becomes mechanical instead of a rewrite: one story becomes a newsletter section and a platform-native post because the channel constraints are already encoded. And you stop being the bottleneck, because the editorial judgment that used to live only in your head now lives somewhere the agent can read it. The profile is not a one-time setup cost; it is the thing that compounds. Every run that reads it is a run you did not have to supervise line by line.
Write it once, edit it as your beat moves
This is the bet Niche is built on. A brand profile you write once is persisted to the desk and threaded through every run, so a story researched on Monday and a LinkedIn post drafted on Friday both read as you without a fresh round of voice instructions. The profile is editable, because voice moves: add a banned word the moment a phrase grates, tighten the framing as your beat sharpens, update channel constraints when you add a surface. The point is not to automate your judgment away. It is to record it once so the work downstream of it can move at the speed of publishing instead of the speed of re-explaining yourself.
What we're tracking next
What we are watching: how often a profile needs editing before it stabilizes, and which field creators skip first. Our working guess is that framing and anti-ICP get skipped most, because voice and banned words feel productive to write, while the harder fields, the ones that say who you are not for and what argument is yours, are the ones that actually stop drift. We will report what the usage data shows once the blog's analytics are live.
Frequently asked questions
What is a brand profile for an AI agent?
A brand profile is a structured document your agent reads before every run. It records the editorial decisions that define your publication: how you sound, the words you use and refuse, who you are writing for and who you are not, what each channel allows, what you cite from, and where readers go next. Instead of re-explaining your voice in every prompt, you write it once and the agent works from it.
How is a brand profile different from a prompt?
A prompt is a single request; it disappears when the context resets. A brand profile is persisted and loaded into every run, so voice survives across tasks and across days. Pasting sample posts into a prompt fixes one draft. A profile fixes the standard.
What should a brand profile include?
Eight fields: voice, lexicon, framing, anti-ICP, channel constraints, proof assets, banned words, and conversion path. Each answers a question an editor asks by reflex. The two fields creators skip most often, framing and anti-ICP, are the two that do the most to stop drift, because they say what argument is yours and who you are not writing for.
How often should I update my brand profile?
Whenever the output tells you to. Add a banned word the first time a phrase grates, tighten the framing as your beat sharpens, update channel constraints when you add a surface. Voice moves, and the profile is meant to be edited, not frozen.
Keep reading