Create a prompt using Prompt management - Amazon Bedrock

Create a prompt using Prompt management

When you create a prompt, you have the following options:

  • Write the prompt message that serves as input for an FM to generate an output.

  • Use double curly braces to include variables (as in {{variable}}) in the prompt message that can be filled in when you call the prompt.

  • Choose a model with which to invoke the prompt or, if you plan to use the prompt with an agent, leave it unspecified. If you choose a model, you can also modify the inference configurations to use. To see inference parameters for different models, see Inference request parameters and response fields for foundation models.

All prompts support the following base inference parameters:

  • maxTokens – The maximum number of tokens to allow in the generated response.

  • stopSequences – A list of stop sequences. A stop sequence is a sequence of characters that causes the model to stop generating the response.

  • temperature – The likelihood of the model selecting higher-probability options while generating a response.

  • topP – The percentage of most-likely candidates that the model considers for the next token.

If a model supports additional inference parameters, you can specify them as additional fields for your prompt. You supply the additional fields in a JSON object. The following example shows how to set top_k, which is available in Anthropic Claude models, but isn't a base inference parameter. 

{
    "top_k": 200
}

For information about model inference parameters, see Inference request parameters and response fields for foundation models.

Setting a base inference parameter as an additional field doesn't override the value that you set in the console.

If the model that you choose for the prompt supports the Converse API (for more information, see Carry out a conversation with the Converse API operations), you can include the following when constructing the prompt:

  • A system prompt to provide instructions or context to the model.

  • Previous prompts (user messages) and model responses (assistant messages) as conversational history for the model to consider when generating a response for the final user message.

  • (If supported by the model) Tools for the model to use when generating the response.

  • (If supported by the model) Use Prompt caching to reduce costs by caching large or frequently used prompts. Depending on the model, you can cache system instructions, tools, and messages (user and assistant). Prompt caching creates a cache checkpoint for the prompt if your total prompt prefix meets the minimum number of tokens that the model requires. When a changed variable is encountered in a prompt, prompt caching creates a new cache checkpoint (if the number of input tokens reaches the minimum that the model requires).

To learn how to create a prompt using Prompt management, choose the tab for your preferred method, and then follow the steps:

Console
To create a prompt

  1. Sign in to the AWS Management Console using an IAM role with Amazon Bedrock permissions, and open the Amazon Bedrock console at Getting Started with the AWS Management Console.

  2. Select Prompt management from the left navigation pane. Then, choose Create prompt.

  3. Provide a name for the prompt and an optional description.

  4. To encrypt your prompt with a customer managed key, select Customize encryption settings (advanced) in the KMS key selection section. If you omit this field, your prompt will be encrypted with an AWS managed key. For more information, see AWS KMS keys.

  5. Choose Create prompt. Your prompt is created and you'll be taken to the Prompt builder for your newly created prompt, where you can configure your prompt.

  6. You can continue to the following procedure to configure your prompt or return to the prompt builder later.

To configure your prompt

  1. If you're not already in the prompt builder, do the following:

    1. Sign in to the AWS Management Console using an IAM role with Amazon Bedrock permissions, and open the Amazon Bedrock console at Getting Started with the AWS Management Console.

    2. Select Prompt management from the left navigation pane. Then, choose a prompt in the Prompts section.

    3. In the Prompt draft section, choose Edit in prompt builder.

  2. Use the Prompt pane to construct the prompt. Enter the prompt in the last User message box. If the model supports the Converse API or the AnthropicClaude Messages API, you can also include a System prompt and previous User messages and Assistant messages for context.

    When you write a prompt, you can include variables in double curly braces (as in {{variable}}). Each variable that you include appears in the Test variables section.

  3. (Optional) You can modify your prompt in the following ways:

    • In the Configurations pane, do the following:

      1. Choose a Generative AI resource for running inference.

        Note

        If you choose an agent, you can only test the prompt in the console. To learn how to test a prompt with an agent in the API, see Test a prompt using Prompt management.

      2. In Inference parameters set the inference parameters that you want to use.

      3. If the model supports reasoning, turn on Reasoning to include the model's reasoning in its response. In Reasoning tokens, you can configure the number of reasoning tokens that the model can use.

      4. In Additional model request fields, choose Configure to specify additional inference parameters, beyond those in Inference parameters.

      5. If the model that you choose supports tools, choose Configure tools to use tools with the prompt.

      6. If the model that you choose supports prompt caching, choose one of the following options (availability varies by model):

        • None – No prompt caching is done.

        • Tools – Only tools in the prompt are cached.

        • Tools, system instructions – Tools and system instructions in the prompt are cached.

        • Tools, system instructions, and messages – Tools, system instructions, and messages (user and assistant) in the prompt are cached.

    • To compare different variants of your prompt, choose Compare variants. You can do the following on the comparison page:

      • To add a variant, choose the plus sign. You can add up to three variants.

      • After you specify the details of a variant, you can specify any Test variables and choose Run to test the output of the variant.

      • To delete a variant, choose the three dots and select Remove from compare.

      • To replace the working draft and leave the comparison mode, choose Save as draft. All the other variants will be deleted.

      • To leave the comparison mode, choose Exit compare mode.

  4. You have the following options when you're finished configuring the prompt:

API

To create a prompt, send a CreatePrompt request with an Agents for Amazon Bedrock build-time endpoint.

The following fields are required:

Field Brief description
name A name for the prompt.
variants A list of different configurations for the prompt (see below).
defaultVariant The name of the default variant.

Each variant in the variants list is a PromptVariant object of the following general structure:

{
        "name": "string",
        # modelId or genAiResource (see below)
        "templateType": "TEXT",
        "templateConfiguration": # see below,
        "inferenceConfiguration": {
            "text": {
                "maxTokens": int,
                "stopSequences": ["string", ...],
                "temperature": float,
                "topP": float
            }
        },
        "additionalModelRequestFields": {
            "key": "value",
            ...
        },
        "metadata": [
            {
                "key": "string",
                "value": "string"
            },
            ...
        ]
}

Fill in the fields as follows:

  • name – Enter a name for the variant.

  • Include one of these fields, depending on the model invocation resource to use:

    • modelId – To specify a foundation model or inference profile to use with the prompt, enter its ARN or ID.

    • genAiResource – To specify an agent, enter its ID or ARN. The value of the genAiResource is a JSON object of the following format:

      {
    "genAiResource": {
    "agent": {
        "agentIdentifier": "string"
    }   
}
      Note

      If you include the genAiResource field, you can only test the prompt in the console. To test a prompt with an agent in the API, you must enter the text of the prompt directly into the inputText field of the InvokeAgent request.

  • templateType – Enter TEXT or CHAT. CHAT is only compatible with models that support the Converse API. If you want to use prompt caching, you must use the CHAT template type.

  • templateConfiguration – The value depends on the template type that you specified:

  • inferenceConfiguration – The text field maps to a PromptModelInferenceConfiguration. This field contains inference parameters that are common to all models. To learn more about inference parameters, see Influence response generation with inference parameters.

  • additionalModelRequestFields – Use this field to specify inference parameters that are specific to the model that you're running inference with. To learn more about model-specific inference parameters, see Inference request parameters and response fields for foundation models.

  • metadata – Metadata to associate with the prompt variant. You can append key-value pairs to the array to tag the prompt variant with metadata.

The following fields are optional:

Field Use case
description To provide a description for the prompt.
clientToken To ensure the API request completes only once. For more information, see Ensuring idempotency.
tags To associate tags with the flow. For more information, see Tagging Amazon Bedrock resources.

The response creates a DRAFT version and returns an ID and ARN that you can use as a prompt identifier for other prompt-related API requests.