Existing brand guidelines

Why does uploading the PDF not work?

Because an uploaded document carries no authority. OpenAI's Model Spec states that file attachments are assumed to contain untrusted data and have no authority by default, and that any instructions inside them must be treated as information rather than as instructions to follow.

So the model can read your guidelines. It will answer questions about them accurately. It will not write to them, because nothing has told it to. The specification also gives the remedy: authority may be delegated by instructions provided in unquoted text, which means one typed sentence saying to follow the rules in that file as though they were written in the instructions themselves.

That is worth trying before anything else on this page, because it is free and it takes a minute. It is also not sufficient on its own, for the reason in the next section.

What survives the conversion from fifty pages to fifteen rules?

Delegating authority to a fifty-page document gets you a model trying to hold fifty pages of context on every request. Anthropic documents what that does: longer files consume more context and reduce adherence, and its guidance is blunt that if a rule keeps being ignored, the file is probably too long and the rule is getting lost.

So the work is extraction rather than upload. Roughly, an agency guide divides into three unequal parts.

Which parts are rules, and which are pictures of rules?

A rule is something two people would apply the same way without discussing it. British spelling. Never more than one exclamation mark per page, or none. Say colleagues, not resources. Second person for the reader, first person plural for us.

A picture of a rule is a sentence that describes the effect a rule would have. We are confident but never arrogant. Warm and human. Approachable yet authoritative. These read well in a deck and two writers will produce different copy from them, which is exactly what happens with a model.

The test is mechanical: could a checker decide whether a sentence breaks it? If yes, it is a rule and it belongs in the file. If no, it is positioning, and it belongs in the document your team reads rather than the one the model does.

How do you extract the rules without rewriting the guide?

• Start from the examples, not the adjectives. Most guides carry a do and do not table somewhere near the back, and it is the densest source of real rules in the document.
• Collect every prohibition verbatim. Anything phrased as never, avoid or do not is already in the right shape and needs no translation.
• Write down the mechanical settings even though they feel too obvious to record. Person, contractions, spelling variant, serial comma, sentence case in headings. They are a handful of characters and they change every sentence you will ever get back.
• Turn each remaining adjective into the behaviour that made somebody choose it. Confident usually means no hedging, so write no hedging words. Warm often means contractions and second person, so write that instead.
• Stop at fifteen. Rank what is left and take the top of the list. Compliance degrades as constraints stack, so the sixteenth rule makes the first fifteen less likely to be followed.

Where the two documents live afterwards, and which one is canonical

The agency guide stays canonical for the organisation. The extracted file is a derivative, and it should say so on its face, with a line naming the source document and its version.

That ordering matters politically as much as technically. Nobody has to accept that a fifteen-line file replaces a signed-off verbal identity, because it does not. It is a rendering of the parts a machine can act on, and the parts it leaves out were never machine-actionable in the first place.

Who owns the file, and what happens when it drifts covers keeping the two in step once both exist, which is the week-two problem.

What to do when the guide and the model's defaults disagree

Write the prohibition explicitly rather than assuming the absence of an instruction is one. A default the model has and your guide does not mention is a default you will get.

Where a rule contradicts something the model does strongly by habit, expect it to need restating in the file rather than inferring. And be aware that if two of your own rules contradict each other, at least one vendor documents that the model may simply pick one arbitrarily, which is covered in what happens when two instruction files disagree.

How to prove to the person who commissioned it that nothing was lost

This is the part that closes it internally, and no vendor writes it because no vendor has to sit in the meeting.

Produce a coverage map: every section of the original, and against each one either the rule it became or the reason it is not machine-actionable. Sections in the second column are not failures. Positioning and audience were never going to become instructions, and saying so explicitly is more convincing than quietly dropping them.

Then show the difference rather than describing it. Run one real task with the file and without it, put the two outputs side by side, and let the person who paid for the fifty pages decide whether the fifteen rules are faithful to them. The verification page sets out that check as a ninety-second procedure with a fixed prompt.

If the answer is that something important is missing, that is a useful result and the fix is one more rule, not a longer file. Rebuilding the file takes about five minutes and nothing you enter is stored or transmitted, which usually settles the other question a brand team asks.

Where these come from

Every claim above is quoted from one of these, and each was read on the date beside it. If one of them has changed since, the page is wrong and we would like to know.

• OpenAI: Model Spec, attachments have no authority by default checked 2026-08-16
• Anthropic: memory, length and adherence checked 2026-08-16
• Mailchimp content style guide, an openly published example checked 2026-08-16