Structured Prompt Builder

Fill in the five parts that actually matter; get a copy-ready prompt.

Press "Build prompt".

The five parts, and why each one earns its tokens

What does not reliably help

Several popular techniques have weak evidence behind them. Threatening or pleading with the model does not change behaviour. Offering tips or promising rewards has no mechanism to act on. Claiming the model is an expert with decades of experience mostly duplicates what a role instruction already does. "Take a deep breath" and similar tokens showed small effects in one narrow benchmark and have not replicated broadly. The parts that do generalise are the ones above: clear task, real context, specified format, stated constraints, and examples when the shape of the answer is hard to describe.

Examples beat description

When you can describe the desired output only awkwardly, show it instead. Two or three input-output pairs — few-shot prompting — usually beat a paragraph of explanation, because the model infers the pattern rather than interpreting your wording. Keep examples representative: if they all share a style or length the model will copy, make sure that is the style you want. Put the most important example last, since recency influences the output more than earlier material.

Order matters

Models attend most strongly to the beginning and the end of a prompt and least to the middle. Put the task and the essential constraints at the start, and repeat the critical instruction at the end rather than burying it among reference material. Long reference documents belong in the middle, ideally trimmed to the relevant passages. This "lost in the middle" effect is why a prompt can contain the right instruction and still ignore it — it was simply placed where attention is weakest.

Iterate on the prompt, not just the output

If the first result is wrong in a specific way, the fix is usually a constraint, not a rephrasing. Too long → add a word cap. Wrong tone → name the audience. Invented facts → require that everything be grounded in the supplied context, and say what to do when it is not. After two or three adjustments, check the size of what you have built with the token counter — prompts grow quietly, and a bloated prompt costs money on every call.