--- title: "AI Prompt Builder: Role, Task & Format | Prompt Pebble" description: "Build clear AI prompts from a role, task, context, constraints, examples and output format, with optional XML tags. Copy, download or save drafts locally." url: https://promptpebble.click/ --- Free · private · runs in your browser # AI prompt builder Fill in the parts of a good prompt (who the assistant should be, the task, the background, your rules, an example and the shape of the answer) and get one clean, well-structured prompt for ChatGPT, Claude, Gemini or any other assistant. ## How the builder puts your prompt together A useful prompt answers the questions a capable new colleague would ask before starting: who am I in this, what exactly do you want, what do I need to know, what must I avoid, what does good look like, and what shape should the answer take? Each box above covers one of them. Leave a box empty and it simply drops out. The role becomes an opening sentence (“You are a careful editor.”), and the rest follows in a fixed order: context, task, constraints, examples, output format. Background comes before the request so the task is read in its light; Anthropic's prompting guide likewise suggests putting long documents near the top and the question after them. Constraints become a bulleted list, one per line, and examples separated by a line of three dashes are numbered. ### XML tags, Markdown headings or plain labels? XML-style tags such as `` and `` make the edges of each part unmistakable, which helps when you paste in documents or code that have headings of their own. Anthropic recommends them for Claude, and OpenAI's GPT-4.1 prompting guide suggests Markdown headings or XML tags as delimiters. Markdown reads naturally in chat apps, and plain labels suit short prompts. All three work with any model, so pick the one you find easiest to edit. Sources: *Prompt engineering overview*, Anthropic documentation; *GPT-4.1 Prompting Guide*, OpenAI Cookbook. ## Quick start 1. Write the task first, in one or two plain sentences. 2. Add a role and whatever background the assistant can't know: your audience, your data, what you've tried. 3. List your rules one per line, then describe the answer you want: length, format, tone. 4. Copy the prompt, download it, or save it as a draft in this browser to reuse later. ## More prompt tools Ready-made templates, exact token counts, batch filling from spreadsheets and a splitter for long documents. - [Prompt templates](https://promptpebble.click/prompt-templates/): 48 original prompts for writing, coding, analysis, learning and business, with variables to fill in. Browse templates - [Token counter](https://promptpebble.click/token-counter/): Exact counts for OpenAI's o200k_base and cl100k_base, a token view and a context-window check. Count tokens - [Prompt variable filler](https://promptpebble.click/prompt-variable-filler/): Fill the {{variables}} of any template from a form or from CSV rows, one prompt per row. Fill a template - [Text chunker](https://promptpebble.click/text-chunker/): Split long text by tokens, characters or words, with overlap, at paragraph or sentence breaks. Split text ## Recommended Two free services that come in handy when you try out AI tools and the links people share about them. Temporary email ### CleanTempMail Trying a new AI app or prompt library that wants an email address before you can look around? CleanTempMail gives you an instant inbox with no sign-up, a random or custom address, and mail that is deleted automatically after 3 days, so your main inbox stays out of their newsletter list. Building an agent that has to click a verification link? Its [Temp Mail API](https://cleantempmail.com/api) is made for automated email testing. [Open a temporary inbox](https://cleantempmail.com/) Disposable cloud browser ### WebTraceCheck Someone shared a link to a “free GPT tool” and you're not sure about it? WebTraceCheck opens it in Brave, Chrome, Edge, Firefox or Chromium on an isolated Linux desktop in the cloud that is deleted after each session, with free daily time and nothing to install. Its free [website audit](https://webtracecheck.com/website-audit) also lists a page's metadata and Open Graph fields. [Try the cloud browser](https://webtracecheck.com/) ## Questions **Which AI assistants does the prompt work with?** Any assistant that takes text: ChatGPT, Claude, Gemini, Copilot, Llama, Mistral and others. The builder only arranges your words. It doesn't call a model or send anything anywhere, so paste the result into whichever assistant you use. **Should I use XML tags or Markdown headings?** Either works. XML-style tags draw a sharp line round each part, which is useful when the context has its own headings, code or long documents. Markdown headings are easier to read and edit by hand. If you're unsure, start with XML tags and switch if the prompt feels cluttered. **Where are my drafts saved?** In your browser's local storage on this device, for this site only. They are never uploaded, and they stay until you delete them here or clear the site's data in your browser. Download a prompt if you want a copy somewhere else. **Why is the token count only a rough figure?** The builder uses a quick rule of thumb, about four characters per token for English, so it stays instant while you type. For an exact count with OpenAI's o200k_base or cl100k_base encoding, paste the prompt into the token counter. **Do I have to fill in every box?** No. Only the task really matters, and every empty box is left out of the prompt. Short requests often need nothing else; a role, some context and an example help most with longer or specialised work.