Answers
You already told it. Twice.
Storing your preferences is the easy half. This is about the other half, where a stored rule actually changes what the agent does.
Last updated
A memory tool saves what you said. That is the easy half, and most products stop there. The hard half is that a saved preference is still text sitting in a window, and the model can skip it.
What changes the outcome is delivery and consequence. The rule arrives with the task it applies to instead of at the top of a long file, and its own checks attach to that task, so skipping it does not pass quietly.
That is the whole difference. Below: why rules get ignored, what you can do about it today for free, and what it looks like when the correction you give once becomes the standard.
It is not just you
"A friend of mine tells Claude to always address him as 'Mr Tinkleberry', he says he can tell Claude is not paying attention to the instructions on Claude.md, when Claude stops calling him 'Mr Tinkleberry' consistently"
Hacker News, December 2025. People invent a canary word so they can tell when their instructions stopped mattering. That is the state of the art.
"Rules are just text, not enforced behavior. I can choose to ignore them."
The model's own answer to a user, quoted in a Cursor forum thread titled "Cursor actively admitting that rules are meaningless, and it doesn't have to follow them". Same thread: "Your carefully crafted rule system is essentially decorative."
Why it happens
Not a bug, and not you writing the rule badly. An instructions file is delivered as context, and context is something the model weighs against everything else in the window. Anthropic writes it in one line:
"Claude treats them as context, not enforced configuration. To block an action regardless of what Claude decides, use a PreToolUse hook instead."
code.claude.com, How Claude remembers your project
Read the second sentence again. The vendor's own advice, when something must actually happen, is to stop relying on the instructions file and put a check in the path.
Three things follow from that, and they are the whole problem:
- Everything loads at once. Your database rule is in the window while the agent works on a button, competing for attention with four hundred lines it does not need.
- Nothing checks. There is no moment where anything asks whether the rule was followed, so following it and skipping it look identical afterwards.
- Nothing notices the drift. You find out when you read the output, which is why people resort to canary words.
The fix, with any tool
Three moves, all free, in order of how much they buy you.
1. Make the rule checkable, not merely clear
"Handle errors properly" cannot be checked by anyone, so nothing can catch it being skipped. "Every API handler returns the standard error shape, and there is a test for the failure path" can be checked in ten seconds. Rewriting your five most-ignored rules this way is the highest-value hour you will spend.
2. Attach the check to the moment, not to the file
A rule in a document is a hope. The same rule written as an item on the task, ticked before the work is called finished, is a step. You can do this with a checklist in your issue tracker. It works because someone has to look.
3. Put a real check in the path where it matters
For the handful of rules you genuinely cannot have broken, take the vendor's advice above: a hook, a lint rule, a pre-commit check. Something that runs. Keep this list short, or you will start bypassing your own guard.
None of that needs a product. The reason people want one is the fourth thing, which you cannot do by hand: making the right rule show up at the right moment, out of hundreds, without you sorting them every morning.
What that looks like as machinery
Mission HQ turns a correction into a standard. There are 14 ways something you say becomes durable, but they all follow the same shape: you say it once, it is stored with who said it, and it comes back attached to the work it governs.
- 1
Correct it once, and the next task already knows
You tell your agent that booking pages must say whether the trip leaves from the port or the beach. One call stores it. Every agent that touches a booking page from that moment on gets it, starting with the very next task, not the next session.
- 2
A fix cannot quietly forget what it learned
When you send something back to be rewritten, that task will not close until the lesson is either saved as a rule or turned down with a reason. The choice is recorded either way, so nobody has to remember to write it down.
- 3
Your yes is the artifact
Approving a way of working stores your actual reply, who you are, and the sample you were shown. Say "make it more formal" and it is read as a correction, not an approval: the draft goes back and comes out again.
- 4
Teach it in your own language
The words you use to describe your work become the routing. Write your rule in Albanian, Spanish or German and it fires before the built-in English patterns, so a project that does not think in English is not a second-class project.
- 5
Approve one page, shape fifty
When you approve one finished example, its full text is stored as the model. Every agent writing a sibling sees the real thing you approved, plus which angles the siblings already used, rather than a description of it.
- 6
Money and login cannot be personalized by accident
Nineteen sensitive areas, payments and authentication and anything destructive, cannot be redirected by a project rule without your recorded approval. It is re-checked every single time, not once at setup.
- 7
Say a law once and it never falls off the page
Some rules govern everything, like every price coming from one place. Those ride at the top of every session and every task in a reserved space that cannot be pushed out to make room for anything else.
- 8
When the advice stops working, it gets examined
If agents keep making a mistake that a rule already warns about, the system opens a diagnosis naming the likely cause: the rule is not reaching them, or it is worded badly, or nothing checks it. It stays in every session start until it is dealt with.
- 9
The same annoyance three times becomes a proposal
Friction reported in three separate sessions is drafted into a written proposal with the evidence attached, waiting for your decision. You are not the one who has to notice the pattern.
Correct it once and watch the next task
14 days, no credit card. Tell your agent one preference, then open the next task and see it arrive.
Where the enforcement actually lands
Precision matters here more than enthusiasm, so here is the shape of it.
The enforcement lives at checkpoints. When a task is handed out, the rule that governs it arrives with it. When the task is closed, the checks that rule brought are items that have to be answered. When a mistake repeats that a rule already warned about, the system opens a diagnosis and puts it in every session start until someone deals with it. Those are the moments where something happens.
The consent record is your own words. When you approve a way of working, what is stored is the reply you actually gave, who gave it, and the sample you were shown when you gave it. That is why "make it more formal" is treated as a correction rather than an approval, and the draft goes back for another pass.
And it is per project, deliberately. What you teach in one client's project governs that project. Nothing leaks sideways into another one, which is the fear anybody running work for several clients actually has.
Questions people ask about this
Why does my AI agent ignore my rules and instructions?
+
How is this different from a memory tool?
+
How do I stop repeating the same instructions to my coding agent?
+
Does my correction apply to my other projects too?
+
What is a canary word and why do people use one?
+
Where does the enforcement actually happen?
+
Two things you can do now
Run the canary test, free, today
Put one harmless line at the top of your instructions file: always address me as Captain. Then work normally. The moment it stops calling you Captain is the moment the rest of that file stopped landing too. Now you know when your rules go quiet, instead of guessing.
Borrowed from the Hacker News comment above. Costs nothing.
Or make the rule arrive with the work
Start a 14-day trial, no credit card. Give one correction, then open the next task and read it back with the correction already inside.
Start free trial