Prompt Builder
Break "what I want it to do" down into structure and generate a prompt you can use right away. A structured prompt usually works far better than a single sentence.
The role decides which part of its knowledge it draws on, the task sets the goal, the context stops it from guessing, the output format makes the result usable as-is, and the constraints block the usual "correct but useless" filler. Any field you skip can stay empty — the Prompt Builder skips it automatically.
When to use it
- You run the same kind of task repeatedly — weekly reports, code review, copy rewrites — and want role, output format and constraints fixed into a reusable template.
- The model keeps returning correct-but-useless generalities, and you want the constraints (what to avoid, what must be included) written into the prompt instead of added verbally each time.
When not to
- The task itself is not defined yet: if you cannot say what the deliverable is or what counts as good, a template only makes a vague request look formal.
- You need few-shot examples, tool-calling schemas, multi-turn state or retrieved documents — those are written by hand or handled by an orchestration framework.
Pure local assembly: nothing uploaded, no model called. Limits: it only concatenates a template, does not judge whether the task makes sense and cannot promise results; at least two fields must be filled; no draft is saved and there is no few-shot, tool definition or multi-turn support.
Frequently asked
Does the builder upload what I type?
No. It is local string assembly: no network request, no model API call, nothing written to localStorage and nothing put in the URL. Close the tab and the input is gone, which also means no draft is saved. The real exposure is the next step — whichever service you paste the finished prompt into applies its own privacy policy and retention rules, so redact confidential material before pasting it anywhere.
Are structured prompts actually better than one sentence?
For tasks with a concrete deliverable, usually yes, because constraints become explicit clauses instead of guesses. But more structure is not always better: persona framing ('you are a world-class expert') adds little on current models and can make answers vaguer. What does the work is a specific task, real input material, a clear output format and instructions for what to do when something is missing. Plan on two or three revision rounds after the first answer, adding the constraints you left implicit — the template is only a starting point.
Does the same prompt work on GPT and Claude?
The structure carries over; the preferences differ. Claude responds well to XML tags and long document context, so wrapping source material in <document> helps; GPT models do better with Markdown sections and a clear system / user split; reasoning models often need no 'think step by step' instruction and can be hindered by it. A prompt getting worse after switching models is normal — rewrite the output-format section in the target model's idiom.