AI most often ignores instructions because the prompt contradicts itself, the rule is buried or phrased as a "don't", the conversation has grown long, or the request carries too many rules at once. The good news is each cause has a specific fix.
This guide gives you seven causes in the order I'd check them (the order is my judgment, so treat it as opinion), a diagnostic table, and the benchmark numbers that explain why long chats go wrong.
Is it really true that AI follows instructions worse over time?
Yes, on published benchmarks, performance falls as a conversation gets longer and as instructions accumulate. Two sources illustrate it:
| Study | What it tested | Result |
|---|---|---|
| Multi-IF (2024) | Multi-turn, multilingual instruction following | All models struggled more as turns increased; o1-preview fell from 0.877 at turn 1 to 0.707 at turn 3 (average accuracy across languages). The authors describe "instruction forgetting". |
| MMMT-IF (2024) | Keeping format instructions across a chat | Average adherence across three models fell from 0.81 at turn 1 to 0.64 at turn 20; human raters followed about 94% of instructions. |
| IFEval (2023) | 25 verifiable instruction types on about 500 prompts | Introduced checkable instructions such as "write in more than 400 words", to measure adherence objectively. |
These are benchmark results on specific older models, not a prediction for your assistant today. Models improve, and results depend on the model and the task. But the pattern is consistent enough to shape your habits: don't assume a rule from message 3 is still in force at message 30.
What are the seven causes and fixes?
Work through them in this order; the first three solve most cases.
| # | Symptom | Likely cause | Fix |
|---|---|---|---|
| 1 | Rule followed once, then dropped | Long chat; instruction forgetting | Start a fresh chat with the rules restated at the top |
| 2 | Does the opposite of a "don't" rule | Negative phrasing puts the unwanted thing in focus | State the wanted behavior ("write in plain paragraphs") |
| 3 | Follows one rule, ignores another | Two rules conflict, or an example contradicts a rule | Remove the conflict; make examples match the rules |
| 4 | Ignores a rule in the middle of a long prompt | Buried instruction | Move key rules near the start or end; use short lists |
| 5 | Follows the letter, not the intent | No reason given | Add one clause of why: "...because readers skim on phones" |
| 6 | Rules fine on small tasks, fail on big ones | Too many tasks at once | Split into steps |
| 7 | Inconsistent between runs | Vague wording; model variability | Make the rule measurable; run twice and compare |
How should I phrase a rule so it sticks?
Say what to do, say it specifically, and say why. Anthropic's prompting guide gives a direct example: instead of "Do not use markdown in your response", try "Your response should be composed of smoothly flowing prose paragraphs." It also notes that giving context or motivation for an instruction helps the model deliver more targeted results. Compare these rewrites (my examples):
| Weak | Stronger |
|---|---|
| "Don't be too wordy." | "Keep sentences under 20 words; one idea per paragraph." |
| "Don't make things up." | "Use only the facts in the notes. If something is missing, write [DATA NEEDED]." |
| "Be creative but professional." | "Open with a concrete scene; keep the rest in plain, professional language." |
| "Never use jargon." | "Use words a high-school student knows; define any technical term in one sentence." |
For measurable rules, the output is easier to check, which you'll need for the testing step below. For formatting rules specifically, how to make Claude follow a strict output format goes deeper.
Does emphasis like CAPS or "CRITICAL" help?
Rarely, and it can backfire with newer models. It's tempting to escalate when a rule is ignored: capital letters, "NEVER", "ABSOLUTELY MUST". Anthropic's guidance for recent Claude models says they are more responsive to the system prompt than earlier ones, so emphasis that once compensated for under-reaction can now cause over-reaction, and recommends softening aggressive language: where you might have written "CRITICAL: You MUST use this tool when...", use normal wording instead. The practical lesson is to use emphasis only after the cheaper fixes (position, phrasing, conflicts) have failed, and to remove it if the model starts applying the rule too widely.
One more case worth knowing: a rule can be followed "too well". If you tell the model "always include an example", it may add examples even where they don't fit. When that happens, add the condition: "include an example when the point is abstract". Rules need scope as well as strength.
What do I do when a long chat has gone off the rails?
Don't argue with it. Start a fresh chat, paste the rules once, and continue from a short summary. A drifted conversation contains your earlier rules, the model's earlier mistakes and your complaints, all competing for attention. A clean restart with a compact brief usually beats five more rounds of "I told you not to".
- Copy the final good version of the output so far.
- Open a new chat and paste your rules block (see the layout in how long should a prompt be).
- Paste the relevant slice of the work and ask for the next step only.
- Reuse the same rules block next time; don't retype it from memory.
This is also why a saved setup helps. Interactive eBook Builder keeps the project's configuration and its editable agent instructions in one place, and the Seed Prompt can be re-run to repair the workspace, so you restart from a known state instead of from a long conversation. If a wrong answer got through anyway, what to do when AI gets it wrong covers recovery.
How do I test whether a fix worked?
Change one thing, run the same input twice, and compare against a checklist. Because outputs vary between runs, a single success can be luck. For a rule like "sentences under 20 words", you can measure it: paste the output into a word-count tool or a short script and count sentences over the limit. Keep the test paragraph and the rule list in a notes file. The method is laid out step by step in how to test and improve a prompt when AI is generic.
A tip (my opinion): if a rule is critical, add a verification step to the end of the prompt, such as "Before answering, check that every rule is met and list any you broke." Models checking themselves isn't foolproof, but it often surfaces misses, and it makes failures visible. When you can, build the check into your workflow: in Interactive eBook Builder, final QA is part of Agent C Polisher's job, so the rule check isn't left to memory.
Before you hit send
- Are all rules phrased as what to do, not only what to avoid?
- Do any two rules or examples contradict each other?
- Are the critical rules at the top or bottom, not buried?
- Is this a fresh chat, or has it run for dozens of turns?
- Can you measure whether each rule was followed?
The takeaway: when AI ignores a rule, assume the prompt or the chat is the problem first. Rephrase positively, remove conflicts, restart long chats and test one change at a time.