Common failure modes
Why does the model guess or wander off my instructions?
When a request is vague, the model fills the gaps with its own guesses; when instructions are buried, it drifts off them. Be specific, state what you want done, and make the key rules easy to find.
Two related failures come from the wording of the request itself. Ambiguity is when you leave a decision open and the model decides for you. Instruction drift is when the model starts to ignore something you told it, usually because the instruction was vague or buried.
Ambiguity is the easier one to picture. “Write something about our product” leaves the length, the audience, and the purpose unstated, so the model picks them. Anthropic’s advice is to be explicit: “If you want ‘above and beyond’ behavior, explicitly request it rather than relying on the model to infer this from vague prompts.” Its rule of thumb is to show the prompt to a colleague with little context, because “if they’d be confused, Claude will be too.”[1] Naming the format, the constraints, and the goal removes the gaps the model would otherwise fill on its own.
A side-by-side comparison makes the gap concrete. “Write something about our product” could come back as a two-sentence tagline or a two-page whitepaper, aimed at an investor or a confused first-time user, and the model has no way to know which you meant. “Write a 150-word landing-page description for a small-business owner who has never used a tool like ours, emphasizing time saved over feature count” removes every one of those open decisions at once. Both are grammatically complete requests; only one of them tells the model what a correct answer looks like.
A subtler form of drift comes from telling the model what not to do. A ban leaves it guessing what you do want. The guidance is to state the positive instead: rather than “Do not use markdown,” try “Your response should be composed of smoothly flowing prose paragraphs.”[1] A positive instruction points at a target; a prohibition only blocks one path.
It also helps to explain why. Adding the reason behind an instruction, “such as explaining to Claude why such behavior is important, can help Claude better understand your goals and deliver more targeted responses.”[1] A model that understands the intent is less likely to drift from it.
Finally, when a prompt mixes instructions, context, and examples, separating them with labels or tags “helps Claude parse complex prompts unambiguously” and “reduces misinterpretation.”[1] Clear structure keeps your rules from getting lost in the surrounding text.
Drift also has a reason to get worse as a conversation runs long, beyond the model simply losing track. Your standing instructions, set in the system prompt or the first message, do not move. The conversation grows around them. What sat right at the start of a short exchange ends up further from the model’s freshest attention as more turns pile on, which is the same edges-over-middle effect the lost in the middle page covers for a single long document, just stretched across turns instead of paragraphs. Restating a rule partway through a long session costs a sentence and resets that clock.
References
- Prompting best practices — Anthropic