"Be more specific" is the most common piece of prompting advice, and it's true and almost useless on its own. It doesn't tell you *how* to be specific in a way that reliably changes output. Below are five concrete techniques, each grounded in documented model behavior rather than general folklore.
1. Give the model a role and a constraint together
A role alone ("act as an expert editor") does less than a role paired with a specific constraint on what to prioritize. "Act as an expert editor. Prioritize cutting redundant phrases over changing tone" gives the model something concrete to optimize for, instead of a vague persona to perform. The constraint is doing more work than the role.
2. Show the model the shape of a good answer, not just the topic
Instead of "write a product description," provide one example of the length, structure, and tone you want, even a rough one, and ask the model to match that shape for a new product. This technique (few-shot prompting) consistently outperforms describing the desired output in the abstract, because "concise but detailed" means something different to every reader, while an example is unambiguous.
3. Separate instructions from content with clear structure
When a prompt mixes instructions and source material in one paragraph, models occasionally lose track of which is which, especially in longer prompts. Structuring a prompt with explicit sections (instructions, then a clearly marked block of source content, then the specific task) reduces this failure mode noticeably. This matters more as prompts get longer and more complex, which is exactly where it's most tempting to skip the structure.
4. Ask for reasoning before the answer, when the task is genuinely hard
This is the technique with the most rigorous research backing of the five: chain-of-thought prompting, covered in depth in our separate explainer, was formally shown to measurably improve accuracy on multi-step reasoning tasks. It's not free. It costs more tokens and time, and for genuinely simple tasks it doesn't help and just adds noise. Reserve it for problems where you'd expect a careful human to also need to think it through rather than answer instantly.
5. Iterate on failure, not just on the first draft
The single highest-leverage habit isn't a prompting technique at all. It's treating a bad first output as diagnostic information rather than a dead end. If a response missed a constraint, the fastest fix is usually to point at exactly what was missed ("this dropped the budget constraint. Keep it under $500") rather than rewriting the entire prompt from scratch. Models respond well to specific, targeted correction; they respond less predictably to a completely different framing of the same request.
An important caveat: prompting technique has real limits
It's worth being honest about what better prompting can't fix. METR's early-2025 randomized controlled trial on experienced open-source developers found that AI coding assistance actually made certain tasks take 19% longer on average, not shorter, a reminder, covered in more depth in our piece on AI coding copilot productivity, that prompting discipline improves the odds of a good interaction but doesn't override a task's genuine difficulty for the tool in question. Good prompting is a multiplier on a tool's real capability, not a substitute for it.
Why this matters beyond convenience
None of these techniques are exotic. They're closer to "how you'd brief a new collaborator clearly" than to any special trick. If you're building anything more involved than a one-off prompt (a workflow that chains several AI steps together) our beginner's guide to building a simple AI agent extends these same principles into a multi-step system.
A caution worth keeping in mind
Better prompting improves the odds of a good output; it doesn't guarantee one, and it's not a substitute for checking the result. See our guide to fact-checking AI-generated content before publishing or acting on anything an AI system produces, however well-prompted.