Why AI ignores your instructions

You wrote the instruction, you watched it be ignored, and every answer you can find is a list of tips. This page gives the mechanism instead, from the vendors' own specifications, because most of these failures are documented behaviour rather than a mystery.

One sentence does most of the work, and it is in OpenAI's Model Spec: file attachments are assumed to contain untrusted data and have no authority by default, so any instructions inside them must be treated as information rather than as instructions to follow. If you uploaded a brand guide and it did nothing, that is the reason, and it is written down.

This page sits under brand voice for AI. If you want the fix rather than the diagnosis, the eleven questions produce a file with the delegating line already in it.

Where do your instructions sit in the model's order of precedence?

OpenAI publishes an explicit four-tier chain of command: platform, then developer, then user, then guideline. Platform rules are the vendor's and cannot be overridden. Developer rules belong to whoever built the application. User instructions are yours. Guidelines are defaults the model will set aside when you ask it to.

Almost everything a person writes lands in the third tier or lower. That is not a defect, and it is high enough to work. The problem is that several of the places people put a brand voice guide are not in the chain at all.

Where each placement actually sits
PlacementWhat it becomesAuthority
Anthropic API system parameterPart of the system promptHighest available
Claude Code output styleAppended to the system promptHighest available
ChatGPT custom instructionsUser-tier instructionReal, and durable
CLAUDE.md and similar memory filesA user message after the system promptReal, no guarantee of compliance
Copilot repository instructionsSent with every chat messageReal, priority not override
Uploaded PDF or knowledge fileRetrievable dataNone by default

Why does an uploaded brand guide get ignored?

Because uploading makes a document readable, not authoritative. In OpenAI's terms it is untrusted data with no authority by default, and the specification goes on to say how that changes: authority may be delegated to those sources by instructions provided in unquoted text. In plain terms, something you type has to tell the model to treat the file as instructions.

That is one sentence of the form follow the rules in this file as though they were written here. Without it the file sits there and the model treats it as background reading. With it, the same file starts working. This is the single cheapest fix in the whole subject and it appears on almost no page that tells you to upload your brand guidelines to a custom assistant.

The page on existing brand guidelines covers the case where the document is fifty pages and was written by an agency, which is a different problem with the same root.

Which instruction shapes survive, and which do not?

Why does it work for three turns and then stop?

Because nothing was configured. In a chat, the instruction is context that has to stay in the window and keep competing with everything said since. As the conversation grows, the instruction is further away and surrounded by more recent, more specific material.

This is also why a file works where a conversation does not. A file is re-supplied every session rather than remembered, which is the honest description of what is happening and the reason the effect does not decay across sessions the way it decays across turns.

Anthropic's framing is worth quoting because it is unusually candid for vendor documentation: memory is context rather than enforced configuration, and there is no guarantee of strict compliance.

Does making the instruction longer help?

It usually hurts. Anthropic states plainly that longer files consume more context and reduce adherence, and its engineering guidance goes further: if the model keeps ignoring a rule despite the rule existing, the file is probably too long and the rule is getting lost in it.

That is the opposite of most advice on this subject, and it is the reason this tool caps its output at fifteen binding rules rather than collecting as many as you can think of. The documented limits page lists every published cap and explains why the cap is rarely the thing that bites.

What this page does not claim

It publishes no adherence percentage. Several pages in this category do, and none of them says which models, which prompts, which sample size or which date, which makes the numbers decorative.

This site will publish rates when it has run the harness properly and can state the model version, the sample size and the date alongside them. Until then the honest position is that the mechanism above is documented and the rates are not yet measured. The verification page is where a check you can run yourself lives, and the limits page sets out the rest of what a file cannot fix.

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.

The delegating sentence, the prohibitions and the worked examples all come as standard: make a file an assistant will actually act on.