> ## Documentation Index > Fetch the complete documentation index at: https://docs.fetchhive.com/llms.txt > Use this file to discover all available pages before exploring further. # AI Prompt > Configure a workflow step that runs an LLM prompt with messages and returns text or structured output Use **AI Prompt** when you want a workflow step to send a prompt to an LLM and store the assistant response for later steps. ## Configuration | Option | Required | Description | | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Name | No | Label for the step in the workflow canvas. | | Charge Type | No | Whether the step uses **Hosted** or **Personal** model billing. | | LLM Model | Yes | The model selected in **LLM Model**. | | Temperature | No | Controls how predictable or creative the response is when the selected model uses temperature. | | Max Thinking Tokens | No | Sets **Max Thinking Tokens** for manual-budget Anthropic Claude and Google Gemini thinking models. Controls how many tokens the model may use for internal reasoning before producing a response. | | Reasoning Effort | No | Sets **Reasoning Effort** (low, medium, or high) for OpenAI and xAI reasoning models. | | Reasoning | No | Toggles reasoning on or off for supported third-party reasoning models such as Cohere. | | Max Tokens | No | Sets the maximum response length for the step. | | Prompt | Yes | Prompt messages in the **Prompt** section. Each message has a role and content. | | Response Format | No | Lets you switch **JSON Schema** on for supported providers. | | JSON Schema | No | Structured response schema used when **JSON Schema** is enabled. | | When the step fails | No | Controls whether the workflow should **Terminate Workflow** or **Continue** if this step fails. | | The **AI Prompt** step uses a wider settings sheet than the other workflow step types. The left side holds the prompt configuration, and the right side shows prompt messages and model responses from test runs in the same view. | | | In **Parameters**, the exact controls change with the selected model: * **Temperature** is shown for standard models. * **Max Thinking Tokens** is shown for manual-budget Anthropic Claude and Google Gemini thinking models. * **Reasoning Effort** is shown for adaptive Claude models such as Claude Sonnet 4.6+, Claude Opus 4.6+, and Claude Fable 5. * **Reasoning Effort** is shown for OpenAI and xAI reasoning models (low / medium / high). * **Reasoning** (enabled/disabled) is shown for supported third-party reasoning models such as Cohere. In **Prompt**, click **Add Message** to build the message list. The first new message is a `system` message. After that, new messages default to `user`. Each message supports variable insertion from the workflow with the **Insert Variable** button. For vision-capable models, `user` messages can also include an image URL. If your selected provider supports structured output, you can enable **JSON Schema** in **Response Format**. This opens a schema editor where you can set a schema name, load an example, and save the JSON schema the model must follow. `assistant` is available as a message role in the editor, but this step still runs as a single prompt execution and stores the final assistant response as the step output. ## Output This step stores the final assistant response as the step output. Use the variable picker to insert the exact reference path for a previous prompt step. In templates and later steps, the base reference is: ```text theme={null} {{STEP_IDENTIFIER.output}} ``` If the response is plain text, reference the full value directly: ```text theme={null} {{step_2.output}} ``` If **JSON Schema** is enabled and the model returns a structured object, you can reference nested fields: ```text theme={null} {{step_2.output.title}} {{step_2.output.summary}} ``` The exact available keys depend on the schema and model response shown in the variable picker. ## Example Add **AI Prompt** from the workflow step picker. Set **Name** to something like `Summarize article`. In **Parameters**, choose your **Charge Type** and **LLM Model**. Adjust **Temperature**, **Reasoning Effort**, **Max Thinking Tokens**, or the **Reasoning** toggle if those controls are shown for your model. In **Prompt**, add a `system` message that explains the task and a `user` message that inserts earlier workflow data such as `{{step_1.output}}`. If you need structured output, enable **JSON Schema** and define fields such as `title` and `summary`. Click **Run** in the step header to test the step. The response appears in the right-hand panel, and later steps can reference that output with the variable picker. {/* Add screenshot: AI Prompt message builder with variable insertion */} {/* Add screenshot: AI Prompt test run showing model response in the right panel */} ## Notes * Use the step identifier shown in the variable picker when you reference this step in later fields. * The **Clear** action removes previous model responses from the test panel. It does not remove your prompt messages. * The step output is the final assistant response, not the full request metadata. * When **JSON Schema** is enabled, later steps can reference the returned object by field instead of parsing raw text. See also: [Creating and Editing](../../creating-and-editing), [Testing and Iteration](../../testing-and-iteration), and [Use your own API Key](../../../workspace/llm-providers)