Improve a prompt
Create a new-and-improved prompt guided by feedback
The prompt tools APIs are in a closed research preview. Request to join the closed research preview.
Before you begin
The prompt tools are a set of APIs to generate and improve prompts. Unlike our other APIs, this is an experimental API: you’ll need to request access, and it doesn’t have the same level of commitment to long-term support as other APIs.
These APIs are similar to what’s available in the Anthropic Workbench, and are intented for use by other prompt engineering platforms and playgrounds.
Getting started with the prompt improver
To use the prompt generation API, you’ll need to:
- Have joined the closed research preview for the prompt tools APIs
- Use the API directly, rather than the SDK
- Add the beta header
prompt-tools-2025-04-02
This API is not available in the SDK
Improve a prompt
Headers
Optional header to specify the beta version(s) you want to use.
To use multiple betas, use a comma separated list like beta1,beta2
or specify the header multiple times for each beta.
Body
The prompt to improve, structured as a list of message
objects.
Each message in the messages
array must:
- Contain only text-only content blocks
- Not include tool calls, images, or prompt caching blocks
As a simple text prompt:
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "Concise recipe for {{food}}"
}
]
}
]
With example interactions to guide improvement:
[
{
"role": "user",
"content": [
{
"type": "text",
"text": "Concise for {{food}}.\n\nexample\mandu: Put the mandu in the air fryer at 380F for 7 minutes."
}
]
}
]
Note that only contiguous user messages with text content are allowed. Assistant prefill is permitted, but other content types will cause validation errors.
Feedback for improving the prompt.
Use this parameter to share specific guidance on what aspects of the prompt should be enhanced or modified.
Example:
{
"messages": [...],
"feedback": "Make the recipes shorter"
}
When not set, the API will improve the prompt using general prompt engineering best practices.
The existing system prompt to incorporate, if any.
{
"system": "You are a professional meal prep chef",
[...]
}
Note that while system prompts typically appear as separate parameters in standard API calls, in the improve_prompt
response, the system content will be incorporated directly into the returned user message.
The model this prompt will be used for. This optional parameter helps us understand which models our prompt tools are being used with, but it doesn't currently affect functionality.
Example:
"claude-3-7-sonnet-20250219"
1 - 256
Response
Contains the result of the prompt improvement process in a list of message
objects.
Includes a user
-role message with the improved prompt text and may optionally include an assistant
-role message with a prefill. These messages follow the standard Messages API format and can be used directly in subsequent API calls.
Currently, the system
field is always returned as an empty string (""). In future iterations, this field may contain generated system prompts.
Directions similar to what would normally be included in a system prompt are included in messages
when improving a prompt.
Usage information
Was this page helpful?