Category
Prompt engineering tools
Write prompts the way the providers’ own guides recommend, see exactly what each API receives, and know the token cost before you send anything. Everything runs in your browser.
Start here
Pick a tool by what you need
Tools
Every tool in this category
- Prompt BuilderBuild prompts with variables and a live token count.
- Prompt DiffCompare two prompt versions side by side.
Coming next: Prompt Template Library.
About
How these tools fit together
A prompt that works in production is mostly structure. The role and rules sit in a system prompt, the material the model should read goes in clearly marked sections, and a few worked examples show the format you want. The prompt builder gives each of those its own box, lets you mark the parts that change with {{variables}}, and shows the result in each provider’s request format, so the same template works with OpenAI, Claude and Gemini.
A prompt that works is worth protecting. When you edit one, the prompt diff shows every word that changed, side by side or inline, with options to ignore spacing and case. It also counts both versions with the tokenizer of the model you call and prices the difference per call and per 1,000 calls, and it lists any {{variable}} or {variable} placeholders that were added, removed or renamed, which is the edit most likely to break the code that fills in the template.
Each API takes the same ideas in a different shape. OpenAI’s Responses API has a top-level instructions field, and its Chat Completions API uses a developer message, which replaces the system message from o1 onwards. Anthropic’s Messages API takes a top-level system with alternating user and assistant turns, and Gemini’s generateContent takes systemInstruction with contents whose roles are user and model. The message format guide compares them side by side.
Every word of a prompt is billed as input tokens on every call, so prompt design is also a cost decision. The token counter measures a prompt with each model’s tokenizer, and the prompt caching calculator shows what a long, fixed system prompt costs once it is cached. The glossary explains system prompts, few-shot prompting and chain of thought in plain terms.
FAQ
Questions people ask
Are these AI prompt generators?
No. They don’t call a model or write prompts for you. The prompt builder arranges what you write into a clear structure and the exact request format for each API, and counts its tokens. Nothing you type is sent to our server.
Should I use XML tags or Markdown in prompts?
Both are recommended by the providers. Anthropic suggests XML tags such as <context> and <instructions> to keep instructions, data and examples apart. OpenAI says Markdown headings and lists can mark sections, and XML tags can show where a document starts and ends. Pick one and use it consistently.
Do the same prompts work across OpenAI, Claude and Gemini?
Mostly. The text can stay the same, but the request shape differs and each model reacts a little differently. Test a prompt on every model you plan to use, and keep it as a template in your code so changes are reviewed like any other change.
How long should a system prompt be?
As long as it needs to be and no longer. It is sent and billed on every call, and rarely used rules compete for the model’s attention. Measure it with the token counter, and if it stays fixed across calls, put it first so prompt caching can reuse it.