Why AI Ignores Parts of Your Prompt
You gave it eight rules and it followed five. The pattern of which three got dropped is usually predictable, and mostly fixable by position.
You wrote eight rules. The output follows five of them. You repeat the three it missed, in capitals, and now it follows those three and drops two others.
This is not randomness and it is not the model being difficult. Instruction dropping has a small number of recognisable causes, and which instruction gets dropped is usually predictable from where it sits and what it conflicts with.
1. It is buried in the middle
The best-documented effect here, and it has a paper. Liu and colleagues at Stanford measured it directly in Lost in the Middle: performance is highest when the relevant information sits at the beginning or the end of the input, and degrades significantly when a model has to reach for something in the middle. That held even for models built specifically for long contexts.
Attention is not uniform across position, and a rule sitting in paragraph four of a nine-paragraph prompt is in the weakest possible spot.
The test costs nothing. Take the instruction being ignored and move it to the last line of your prompt. If compliance jumps, you had a position problem rather than a wording problem.
The general fix is structural: put the non-negotiable constraints at the end, immediately before the model starts generating. Reference material goes in the middle, where it is being retrieved from rather than obeyed.
2. Two of your rules disagree
Very common in prompts that grew over time, because each rule was added to fix a specific failure and nobody re-read the whole thing afterwards.
Be concise. Two sentences maximum.
...
Explain your reasoning step by step before answering.Both are reasonable. Together they are impossible, so the model satisfies one and the other looks ignored. Softer versions of this hide well: "professional tone" alongside "sound like a real person", or "use only the provided context" alongside "explain any unfamiliar terms".
Read your prompt as an instruction list with the content stripped out. Contradictions are obvious in that form and nearly invisible in prose.
3. It is phrased as a prohibition
Negative instructions perform worse than positive ones, consistently enough to plan around. "Do not use bullet points" underperforms "write in continuous prose". "Avoid technical jargon" underperforms "explain at the level of a smart fifteen year old".
A prohibition tells the model what the target excludes. A positive instruction tells it what the target is, which is a much easier thing to aim at. This is a big enough effect on its own that it has its own explanation.
4. The prompt is simply too long
There is a point past which adding instructions reduces total compliance rather than increasing it. Every rule competes for the same finite attention, and the twentieth rule does not arrive for free.
Symptoms are distinctive: compliance that was fine at six rules degrades across the board at fifteen, and different rules get dropped on different runs rather than the same one every time. That variability is the signature of overload rather than of any single bad instruction.
The fix is not better wording. It is fewer instructions per call, which usually means splitting the work into a chain where each step carries only the rules it needs. Worth checking your prompt against how long a prompt should actually be before assuming length is free.
5. It has been diluted by conversation
The instruction was followed in the first three turns and stopped being followed by turn twelve. Nothing about the instruction changed. Everything around it did.
As a conversation grows, your original instructions become a smaller and smaller fraction of the input, competing with thousands of tokens of intervening exchange. The effect has a name, context rot, and long agent sessions are where it shows up most severely.
Two practical responses: restate critical constraints at the point of use rather than only at the start, and treat a degrading conversation as a signal to start a fresh one with a clean summary rather than something to push through.
6. It was never actually an instruction
The most common cause in prompts written by people who write well.
It would be great if the summary could focus on the financial implications.That is a preference expressed politely. Compare:
Summarise only the financial implications. Omit everything else.Hedged phrasing gets treated as a hint that competes with everything else in the prompt, rather than as a constraint. Scan your prompt for "try to", "if possible", "ideally", "it would be good if", and rewrite each one as a direct statement.
A diagnostic order that works
When something is being ignored, work through this rather than rewriting blindly:
Move it to the end. If that fixes it, stop. Position problem.
List every instruction on its own line. Look for contradictions. Remove one side.
Rewrite prohibitions as positive statements. Say what to do, not what to avoid.
Count the rules. More than about eight in one call, split the task.
Check the conversation length. Long thread, start fresh with a summary.
Remove hedging. Every soft request becomes a direct one.
Most cases resolve at step one or step six, which is a useful thing to know before spending an afternoon on step four.
What does not help
Capital letters and emphasis. IMPORTANT and bold markers produce a small effect at best, and once several instructions are marked important, none of them are. Position is a stronger lever than emphasis, and it is free.
Repeating the instruction three times. This adds length, which is one of the causes you are trying to eliminate. Repeating once at the end is different and does help, because that is a position fix rather than a volume fix.
Adding a rule to fix each failure. This is how prompts reach fifteen contradictory instructions. When something fails, first ask which existing rule is competing with it.
Questions people ask
Is this different across models?
The causes are the same. Sensitivity differs, and prompts tuned against one model often lose compliance when moved, which is a reason to write for portability when you have the choice.
Does putting instructions in a system prompt help?
Usually yes. System prompts are weighted more heavily for persistent behavioural rules, and it separates the durable constraints from the per-request content. It does not exempt them from the conflict or length problems above.
Why did it work yesterday and not today?
Check whether the conversation is longer, the input document is larger, or the provider updated the model. All three change the balance without changing your text. The systematic version of this is debugging a prompt that stopped working. The broader foundations are in our prompt engineering guide.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
