Why “YOU MUST” makes your prompt worse

Almost everyone who has written a prompt has done this. The model ignored an instruction, so you went back and made the instruction louder:

~~Use the search tool when the question needs current information.~~ CRITICAL: You MUST ALWAYS use the search tool. Do not answer from memory under any circumstances.

It feels like the right move. It is how you would escalate with a person who was not listening. And on older models it often worked, because they genuinely were reluctant — they would hedge, refuse, or quietly skip instructions, and turning up the volume pushed them over the line.

That reluctance is gone, and the technique that fixed it now causes the opposite problem.

What changed

Current models follow instructions literally. They do not weigh how emphatic you were and then decide how much to comply — they take the instruction at face value and apply it. When you write "ALWAYS use the search tool", the model does exactly that: it searches when you ask what two plus two is.

This is called overtriggering, and it is the dominant failure mode of aggressive prompting today. The tool fires when it should not. The rule applies where it does not belong. The disclaimer appears on every answer. You did not get more compliance — you got compliance in places you never wanted it.

The cruel part is that overtriggering looks like a different problem. The model starts behaving strangely, so the natural instinct is to add another rule to rein it in, phrased just as forcefully. Now you have two aggressive instructions fighting each other, and the prompt gets worse with every round of "fixing".

What it looks like

Here is a prompt with the problem. It is not a strawman — it is close to what people actually write when a model has disappointed them once:

CRITICAL: You must summarize this report in 3 bullets for our exec audience. Ensure you NEVER MISS anything important.

Run it through the checker and the phrasing check fails. That is not a stylistic complaint — it is a prediction about behaviour:

Two separate things are wrong here, and they are worth separating.

"CRITICAL:" and "You must" do nothing. They add tokens and they add pressure, but they do not add information. The model already knows you want a summary — you asked for one. Emphasis is not an instruction.

"NEVER MISS anything important" is actively harmful. Read it literally, as the model will: nothing may be omitted. But the entire job of a summary is omission. You have written an instruction that contradicts the task, and then shouted it. Expect a "summary" that reproduces most of the report, because that is the only way to satisfy what you actually said.

The fix is quieter, not louder

Replace emphasis with specificity. Every time you want to write "you must", ask what you are actually afraid the model will get wrong, and then say that instead.

Summarize the report below in 3 bullets, each under 20 words, for an executive audience.

### Requirements
- Cover the revenue change, the cause, and the recommended action.
- Leave out anything that does not change a decision.

### Report
"""
Q3 revenue fell 8% against forecast, driven by churn in the SMB segment.
"""

Look at what replaced the shouting. "NEVER MISS anything important" became a list of the three things that actually matter, plus an explicit instruction to drop everything else. That is a bar the model can check its own work against. "Do not miss anything" is not.

The data also moved inside delimiters, which is a separate rule, but it belongs to the same instinct: be precise about the boundaries rather than emphatic about the intent.

The rules of thumb

Never write an instruction in capitals. If it matters, it matters in lower case. Capitals are how you tell the model to overtrigger.

Delete "CRITICAL:", "IMPORTANT:", "You must", "at all costs", and "under no circumstances". Read the prompt without them. It almost always says the same thing, because those phrases carry no information — they carry a mood.

Replace "always" and "never" with a condition. "Always use the search tool" becomes "use the search tool when the answer depends on information after your training cutoff". The first is a demand; the second is a rule the model can apply correctly.

If the model ignored an instruction, the instruction was unclear, not too quiet. This is the one that takes discipline. The reflex to escalate is strong and it is almost always wrong. Go back and ask whether the instruction was checkable. "Be concise" is not checkable. "Under 100 words" is.

Why this is counterintuitive

Every other domain rewards emphasis. Bold text in an email, a raised voice in a meeting, "URGENT" in a subject line — these all work on people, because people infer intent from tone and adjust.

Language models do not infer intent from tone. They read what you wrote and do it. The whole basis of the escalation instinct — if I say it more forcefully, they will take it more seriously — has no mechanism to act on.

So the technique that feels most like taking control is the one that quietly hands it away.

Every check in this article ran live in your browser. The same nine run on any prompt you paste in, for free.