A brand context file for Claude: what it fixes, what it misses

You ask how last week went. Back comes a tidy report that calls your booked calls “purchases”, treats a lead form as the end of the funnel, and analyses ad sets when you wanted ads. Nothing in it is invented. It is all built on the account. It is just built on an account read by something that does not know what you sell.

The usual advice is to write a brand file so the assistant “has more context”. That is true and it is not the interesting part. The interesting part is that Meta’s connector reads your vocabulary as routing information, and the tool schemas say so in writing.

Your words are a routing key

Nearly every tool on the official connector carries a parameter called advertiser_request. Read on 2026-09-12, its description opens by telling the assistant to record what the advertiser is actually asking for, in their exact words — and then goes further, asking for it quoted from their own messages word for word wherever you can.

Then it forbids the tidying-up a model does by reflex. No paraphrasing, no shift into a more formal register, and specifically do not upgrade their plain words into domain or industry terms or add metric abbreviations or technical, product, or system field names they did not say themselves. Your phrasing is not normalised on the way to Meta. It is preserved on purpose.

The second place is blunter. The performance-trend tool’s analysis_level parameter is chosen by a written keyword rule, not by judgement. One branch reads: query contains “ad” or “ads” as a separate word, make one call at ad level. The schema then spells out what does not count as that word — adset, ad set, ad sets, advertisement, advertiser. So “how are my ads doing” and “how is my advertising doing” do not land in the same place, and if you call your ad sets “campaigns” in conversation, the branch above fires and you get a campaign-level answer to an ad-level question.

Getting the wording right is a habit, not a prompt you paste once. The Claude Ads Operator course teaches it as part of the weekly routine, on top of an account that is properly connected.

The third place is where the guessing actually happens. Two further parameters, conversation_intent and conversation_topic, ask the assistant to classify you. Intent is your “primary goal or business objective” and must be one of seven fixed values: acquire new customers, increase awareness, increase engagement, increase value, increase volume of results, optimize cost outcomes, optimize delivery. Topic is the “primary problem area” and must be one of ten, from auction and delivery through to signals. The schema tells the assistant to provide the parameter even if the query is indirect or phrased as a question about metrics.

So whatever you are trying to do this month gets mapped onto one of seven words before Meta sees it, by a model working from whatever you happened to type. “Stop wasting spend on the wrong leads” is at least three of those values, depending on which clause it latches onto. Naming the value yourself removes the guess. Nothing else you write does.

What the file controls, and what it does not

What the guides get wrong about it

The vendor blog Adspirer, in a guide updated 24 August 2026, describes the context file as a safety mechanism: your file’s safety rules are “your protection layer. They are not suggestions — they are constraints the agent enforces before taking any action.” The first half is a fair description of intent. The second half is not how any of this works. A line in a file is a sentence in the prompt, read by the same model that is deciding what to do, and it competes with everything else in the conversation.

The constraints that hold are the ones living outside the model. Campaigns and ads created through the official connector come back paused, and the creation tools say so in their own descriptions, which is why drafting is safe in a way that instructions are not. Above that sit Meta’s portfolio-level server rules. The sort worth keeping in your head is which of your rules would still hold if the model ignored the file entirely.

1. The words you use for your own things highest leverage

   What you call a conversion, what you call a campaign versus an ad set, the names of your products and offers, and the plain-English phrase you use for the thing you actually sell. This is what the verbatim-capture instruction is preserving, and it is what the level rules read.

   The trade: it only works if you then use those words yourself. A file listing vocabulary you never type changes nothing, because the parameter copies what you said, not what your file said.

2. Your objective, in the connector’s vocabulary removes a guess

   Pick the value from the list above that genuinely matches what this quarter is for, and write it down. Otherwise it is selected for you, per call, from context.

3. The facts a report will otherwise fill in plain prose

   What happens after the click. Whether a lead is the outcome or a step. Which metric you judge on and which you ignore. Anything seasonal that makes a bad week normal. None of this is visible in the account, so none of it reaches Meta unless it reaches the model first.

A context file is an input to the assistant. It is never a permission system.

• Writing rules and calling them guardrails. “Never change budgets” in a file is a request. Budget changes stay a human decision because you keep them one, not because the file says so.
• Putting performance figures in it. Anything numeric in the file gets quoted back as current long after it stopped being true, with the confidence of something you wrote yourself.
• Writing it in jargon you never speak. The file that helps is the one matching how you actually type at eleven at night, because your typing is what gets copied into the call.
• Expecting it to fix the arithmetic. Better context produces a better-aimed report, not a more accurate one — the figures still need checking.

Try this tonight Before asking for any data, have the assistant tell you back what your business sells and what counts as a conversion for you. Everything wrong or hedged in the paragraph that comes back is a line for the file, and everything it had to guess at is a place the connector would have guessed too. That is the first draft, written in one question.