Why a checklist catches drift and taste does not

The people whose feeds drift are not the people with bad taste. Very often they are the people with the best taste in the room, which is the part that makes this worth writing about.

Here is the asymmetry, and everything else follows from it:

Taste is superb at judging one piece of work. It is structurally unable to judge a hundred.

A good editor reads a post and knows whether it is good. What they cannot do — what nobody can do — is hold the previous ninety-nine in working memory while they read the hundredth. So each piece is evaluated against an impression of the brand rather than against the brand, and impressions update. Slowly, in the direction of whatever you read last.

That is the mechanism of brand drift, stated as precisely as we can: it is not carelessness, it is that the evaluation function has no memory and the archive does.

Why every individual post passes

Drift has a signature that makes it nearly invisible from the inside. Every single piece is defensible on its own.

The post that was a bit more casual was casual for a reason — the topic was lighter. The one that opened with a question opened that way because the hook needed it. The one that used "we" instead of "you" was announcing something. Each decision was locally correct, made by somebody competent, for a stated reason.

Twenty of them and the archive has no single author.

You cannot catch this by reviewing pieces harder. The review was not wrong. It was answering a different question from the one that matters, and no amount of diligence converts one into the other.

What a checklist actually does

A checklist is not a substitute for judgement and anyone who has worked with a bad one knows it. What it does is carry the state that judgement cannot.

Four questions, run against each piece before it goes out:

1. Does it break any of our refusals? Not "is it good" — does it do a specific thing we decided we do not do.
2. Is the address consistent? "You", "we", or neither. Pick one; this is the most detectable inconsistency and the most commonly broken by accident.
3. Is the sentence rhythm in range? Long-and-then-short, or steady medium. Not a rule about length — a rule about the pattern.
4. With the logo removed, would a regular reader attribute this to us?

Question 4 is the real one. Questions 1 to 3 are how you answer it without arguing about taste, because they are checkable and taste is not.

The reason this works is not that the checklist is smart. It is that the checklist is the same every time, which is exactly the property an impression does not have. It is externalised memory for a judgement process that has none.

The failure mode of checklists, honestly

Checklists get abandoned, and usually for one of three reasons worth naming in advance.

Too long. A twelve-item checklist gets run properly twice and then eyeballed. Four items get run. There is a real trade here and the short one wins, because a checklist that is not run has no value at all regardless of its coverage.

Not falsifiable. "Is it on brand?" is not a checklist item, it is the question the checklist is supposed to answer. Every item must be one somebody can fail and know they failed.

Run by the writer only. Self-review catches item 1 and misses item 4 reliably, because the writer is the one person in the world who cannot read their own draft as an outsider. Item 4 needs a second pair of eyes, or a gap of a day, or something mechanical.

Where a machine actually helps, and where it does not

This is the part most tools oversell, so precisely:

A machine is good at the checkable half. Items 1 to 3 are mechanical. Did the draft use a banned phrase, did the address switch, does the rhythm sit inside the range the rest of the archive occupies — those are measurable across a body of work, which is the exact thing a human reviewer cannot hold.

A machine is not good at item 4. "Would a reader attribute this to us" is a judgement about a relationship, and no scoring system has that. Anything claiming to automate it is scoring a proxy and calling it the thing.

Our brand consistency checker does the first half and not the second, and it has a further limitation worth stating: it reads writing and palette contrast, so a feed that drifted through imagery — inconsistent crops, borrowed stock, changing photographic treatment — will look fine to it. That is a real hole and it is the most likely way it tells you something reassuring that is not true.

The design intent is that the machine handles what accumulates and the human handles what matters, rather than the machine pretending to do both. The same split is why the on-brand content checklist keeps some items as a document rather than moving all of them into the product: a document asks you to be diligent at exactly the moment your diligence has run out, and the answer is to move the mechanical half out and leave the judgement half in.

If you only do one thing

Do not start with the checklist. Start with the refusals.

Item 1 is worthless until somebody has written down what this brand does not say, and that is the step nearly everyone skips — it is the one that costs something, because a refusal excludes work you might otherwise have shipped. Six branding words used to avoid a decision is about the vocabulary that makes skipping it feel like progress, and the twelve voice cards are worked examples, each printing a refusal column beside its "says" column.

Then read your last twenty posts in one sitting, in order. That reading is the smallest useful brand audit there is. Not to judge them — to see whether they have one author. That reading is the only reliable drift detector anyone has, it takes about an hour, and almost nobody does it on purpose.

A checklist per piece is the right shape for regular posting and the wrong shape for an announcement, where the rule changes with the occasion: a price increase and a new hire fail in opposite directions. How to announce it takes fifteen moments one at a time, and the thirty drafts behind it are the measurement that decided which advice was worth giving.

Two adjacent arguments this one leans on. The checklist only works if somebody wrote the refusals down, and the style guide nobody opens is about why the document holding them is usually too long to be read at the moment it is needed. And the reading procedure at the end of this post is the same test for whether a brand has outgrown being a look.