Prompt builder with variables, live preview and token count

A good prompt is mostly a matter of putting the right things in the right place. The builder gives each part its own box and assembles them in an order drawn from the providers’ own prompting guides, checked on 2026-10-12. It doesn’t use a model to write or “improve” anything: what you see is exactly what you typed, arranged and labelled.

How the prompt is assembled

With a system prompt switched on, the role comes first, followed by your constraints, the output format and any examples. OpenAI’s prompt engineering guide suggests the same order for a developer message: identity, then instructions, then examples. The user message holds the context first and the task last. That follows Anthropic’s long-context advice to put documents and data near the top and the question at the end, which it says can improve response quality by up to 30 percent in tests with complex, multi-document inputs. With the system prompt off, everything goes into one user message, still with the task last.

Sections are marked with XML tags by default, such as <context> and <output_format>. Anthropic recommends XML tags so Claude can tell instructions, context, examples and inputs apart, and asks for consistent, descriptive tag names. OpenAI says Markdown headings and XML tags both help mark sections, so you can switch to Markdown headings or plain labels.

Variables

Anything written as {{name}} becomes a variable with its own box. Names start with a letter or underscore and can contain letters, digits, underscores and hyphens. Values are inserted in a single pass, so a value that itself contains {{…}} is left as written rather than expanded again. A variable you haven’t filled stays visible as a placeholder, and the preview warns you before you copy it.

Examples: in the prompt or as turns

Examples are one of the most reliable ways to fix a format, and Anthropic suggests 3 to 5 varied ones in <example> tags. You can also send them as earlier conversation turns: a user message with the example input, then an assistant message (a model turn in Gemini) with the output you want. The final message is always from the user. That matters for Claude: from the Claude 4.6 models onwards, a request ending in a partial assistant message (a prefill) returns an error.

Each provider’s request format

The same prompt is shaped differently for each API. OpenAI recommends the Responses API for new projects, where the system prompt is the top-level instructions field; Chat Completions is still supported and takes it as a developer message, which replaces system from o1 onwards. Anthropic’s Messages API has a top-level system and requires max_tokens. Gemini’s generateContent takes systemInstruction and contents, whose roles are user and model. Google now recommends its newer Interactions API for new work but says generateContent remains fully supported. The message format guide covers images and tool calls too, and the model ID reference lists the IDs the snippets use.

Token count and cost

The token count runs in your browser with the chosen model’s own tokenizer where one is public: OpenAI’s encodings and open-weight models’ tokenizer files. Claude and Gemini tokenizers aren’t published, so for those the builder shows a range from OpenAI’s count, the same method as the context window checker; the token counter can fetch the exact figure with your own key. The cost per call adds your expected reply at the model’s output price, using the same prices as the cost calculator. The count covers the prompt text; each API adds a few tokens of formatting per message. If the role, rules and examples stay the same on every call, they sit at the start of the request, which is where prompt caching needs them to be.