Five prompting habits that Anthropic and OpenAI both put in their own documentation

Most prompting advice circulating online is folklore: a phrase somebody pasted in once, got a good answer, and passed on. The two labs that build the models most people use publish their own guidance, and it reads nothing like that. It is unglamorous, it is specific, and — written for different models, different years, different audiences — it keeps arriving at the same handful of moves.

Here are five that appear in both, with the reason each one works.

1. Write the instruction as a command, not as a question

"Can you suggest changes to this function?" and "Change this function so it handles an empty list" are not the same request. The first asks whether suggestions are possible; a literal reader answers that question and stops. The second asks for the work.

Anthropic's guidance is explicit that current Claude models take instructions literally and do what is asked and not more. OpenAI's cookbook makes the same point from the other direction: it treats ambiguity as a defect and tells the model, when a request is unclear, to state its best-guess interpretation and proceed rather than stall.

The pattern: name the verb, name the object, name the boundary.

2. Give explicit permission to say "I do not know"

This is the cheapest reliability fix available and almost nobody uses it.

A model pushed to answer confidently on something it has no information about will sometimes fill the gap. Anthropic's documentation addresses this directly: adding explicit permission to acknowledge uncertainty prevents it. One sentence in your prompt — if you do not have enough information, say so instead of guessing — changes the outcome, because you have removed the implicit instruction that an answer is mandatory.

It is worth adding to any prompt where a wrong answer costs more than no answer: anything involving figures, citations, legal or medical specifics, or a claim you plan to repeat to somebody else.

3. Put the context above the question, not below it

Both guides order the prompt the same way: material first, instruction last.

Anthropic's practical framing is that long inputs should be structured clearly, with the critical detail at the start or the end rather than buried in the middle. Newer models handle long context far better than the generation that produced the original "lost in the middle" research, but the ordering habit costs nothing and still helps — and it makes your own prompt easier to edit, because the ask stays in one place at the bottom instead of being hidden three paragraphs into a wall of pasted text.

4. One prompt, one task

The temptation with a large context window is to load everything in and ask for six things at once. Both guides push back.

Anthropic's version of the argument is the interesting one, because it explicitly separates the reason from the old constraint: breaking a large job into smaller chunks is worth doing not because of context limits, but because a focused task with clear boundaries produces higher quality than a prompt chasing several objectives at once. That reframes splitting from a workaround into a technique — it stays useful no matter how large windows get.

If you want the summary, the risk list and the email draft, that is three prompts. You will also be able to tell which one went wrong.

5. Decide how proactive the model should be

This is the habit most people have never made a decision about, because it is invisible until you look for it.

OpenAI's GPT-5 guide devotes a section to calibrating "agentic eagerness" — how far the model runs on its own before checking back — and notes that lowering reasoning effort narrows the scope of that behaviour, cutting tangential tool calls and latency. The default is thorough context-gathering, which is right for an open-ended investigation and wrong for a small, well-specified edit.

So decide, per task, which one you are running. A well-defined change wants a tight leash and a low effort setting. A vague investigation wants the opposite. Accepting the default silently means half your tasks get the wrong posture.

The pattern underneath

Read the five together and they are one instruction wearing five hats: say what you actually want.

Command instead of question removes ambiguity about whether you want the work done. Permission to decline removes the implied requirement to produce something. Context first removes ambiguity about what the material is for. One task removes ambiguity about which goal wins. Setting eagerness removes ambiguity about how far to go alone.

That is why the advice converges across two labs and several model generations. The models keep changing. The thing being fixed — an underspecified request — does not.

Check the date. Both vendors revise these documents as models ship. The links under Sources are the primary versions; if you are reading this months later, read them rather than us.