Skip to content
Early access is open
Repic
LoginJoin early access
← Back to the journal
Playbook

Describing your brand once, instead of explaining it every time

Prompting per post is not lazy — each prompt is a reasonable summary written under time pressure. That is exactly why the archive drifts. The argument for holding a brand as state rather than re-describing it, and the three cases where prompting is the better answer.

By Saffi & JaveriaAugust 17, 2026 · 6 min read

The case against prompting per post is usually made as an efficiency argument: you retype the same thing every time, so store it once. That argument is true, small, and not the interesting one.

The interesting one is about variance.

Every prompt is a lossy summary, and no two agree

When you describe your brand at the top of a prompt, you are compressing something large into a sentence or two, in a hurry, aimed at the post in front of you.

Monday: "friendly, expert, for solo founders."
Thursday: "approachable but authoritative, small business audience."
The following week: "straight-talking, no fluff, for people who are time-poor."

Every one of those is a reasonable description of the same brand. None is wrong. And they will produce measurably different writing, because the model is doing exactly what it was asked and the asks differ.

The drift is not in the model. It is in the compression. You are re-summarising from memory each time, and human memory of an abstraction is not stable across weeks. The output faithfully reflects an input that moved.

This is the same mechanism as brand drift in human-written work, arriving faster because generation is faster. Speed does not cause it; it just removes the interval in which somebody might have noticed.

What "state" buys you, precisely

A brand persona is a structured description held once and read by every engine automatically. Three consequences, and the third is the one people do not expect:

1. The description stops varying. Same input, so variance in the output is variance in the request rather than in the brand. That is the whole point and it is the least interesting of the three.

2. It can be deeper than you would ever type. Nobody writes a paragraph about who they are not for at the top of a prompt about a carousel. But that field is exactly what stops the model reaching for the generic option — and it only exists if it is stored somewhere other than the prompt.

3. It can be scored, and it can go stale. This is the one that matters. If the description is a field in a structure rather than a sentence in a prompt, something can look at it and say this field is thin, or you changed your positioning and forty pieces of work now predate it. A prompt cannot be audited. It is gone the moment it runs.

That last point is why our quality engine exists and it is the honest reason to hold a brand as state at all. The value is not the storage. It is that stored things can be inspected and prompts cannot.

The thin-field problem

There is a failure that looks like a model problem and is not, and it is worth naming because it is the single most common cause of "the AI writes generic stuff".

A field you left blank is not neutral. It is a slot the model fills with the statistically most ordinary option.

Leave the audience thin and it writes for everybody, which is a real audience — the average of all of them. Leave the position undecided and it produces the average position. Every unanswered question gets answered by the distribution, and the distribution is by construction the most generic answer available.

This is why depth in the description is not perfectionism. Each field left thin is a decision handed to a model that has no information about you, and it will make it in the most average way possible, every single time, consistently. You will get consistency. It will just be consistency with everyone else.

When prompting per post is the right answer

Holding state is not always correct, and three cases where it is not:

One-offs. A single asset, once, for something you will not repeat — a prompt is faster and the state has nothing to amortise over. This is what our free tools are for and why they are stateless by design.

You have not decided yet. If the brand genuinely is not settled, encoding an undecided brand as state makes a guess authoritative. Prompt while you experiment; the variance is a feature at that stage, because you are looking for what fits. Encode it once something has been chosen.

Deliberate departures. A post that intentionally breaks register — an apology, an announcement, something serious in a light brand — should not be forced through the standard description. State should be the default, not a cage, and any system that cannot be overridden for one piece is worse than no system.

The test

If you want to know whether this matters for you, the check takes an hour and needs no tool:

Open the last twenty things you published. Read them in order. Ask whether they read as one author.

If yes, whatever you are doing works and you can stop reading. If no, ask whether the individual pieces were bad — they almost certainly were not — and if they were all fine, the problem is not craft. It is that the description behind them moved, and no amount of writing better fixes an input that varies.

The fix is to write it down once, properly, including the parts that cost you something: who you are not for, what you refuse to say. Six branding words used to avoid a decision covers why those two are the ones people skip. The rest is bookkeeping, and bookkeeping is what software is good at.

If you want the mechanics rather than the argument, the branding engine covers how the persona is built and read, why prompt-per-post content fails makes the failure case in more detail, and AI personal branding is the same argument for a person rather than a company.

One caveat this post does not cover: a stored description is only as trustworthy as the claims made about it. Nine questions to ask before anything clones your voice is the same scepticism applied to the part of the product that handles biometric data, and it names four questions we do not yet pass.

One measured illustration of the same mechanism, if you want evidence rather than argument: thirty announcement drafts, read by the same engine, where first-person mentions fell on fifteen of fifteen while sentence length barely moved. The variable people reach for is not the variable that separates the drafts.

On the field that most often stays thin: content pillars are not content categories is about the one where a topic bucket gets written where an argument was needed.