LLM Prompt
The workhorse of every AI-based automation. Send a prompt to an AI model and capture the response.
What it does
Sends one request to an AI model — a system instruction plus a user message — and writes the response to memory. Other steps can then read the response, send it on, store it, or use it as input for another model call.
Use this step whenever you want the AI to do something straightforward: summarize, translate, classify, rewrite, draft, extract.
What you configure
| Field | What it controls | Required | Notes |
|---|---|---|---|
| Persona | A predefined behavior description (tone, role, rules). | optional | Choose from the workspace's personas, or leave empty. |
| Model | Which AI model executes the request. | required | Pick from the workspace's connected models. See Models. |
| System instruction | The brief for the AI: who it is, what it should do, what rules apply. | recommended | Plain language. The more specific, the more reliable the output. |
| Context (memory input) | Memory keys whose values are made available to the prompt. | optional | Use this to provide background data — for example "the customer record from the previous step". |
| LLM options | Provider-specific parameters (temperature, max tokens). | optional | Defaults are sensible. Tune only when you know what you want. |
| Request (memory input) | The actual question or task. Typically pulled from memory (the trigger's payload, the previous step's output). | required | This is the "user message" the AI sees. |
| Output format | Plain text, JSON, or a JSON shape you describe. | optional | Use structured when downstream steps need a specific shape. |
| Result text (memory output) | The text content of the AI's response. | required | This is what you usually want to feed into the next step. |
| Raw result (memory output) | The full response object including metadata (usage, model info). | optional | Useful for logging, less so for chaining steps. |
📷 SCREENSHOT: The LLM Prompt configuration panel with model selector, system instruction, and request fields visible.
Example scenario
Daily ticket summary. A workflow fetches yesterday's support tickets, passes them through an LLM Prompt step with a system instruction like "You are a support analyst. Group tickets by priority and write a one-paragraph summary.", and emails the result to the team lead.
Recommendations
- ✅ Put style guidance ("Use bullet points. Maximum 200 words.") in the system instruction, not the request. It is easier to maintain there.
- ✅ Use structured output when a later step needs to read specific fields. Plain-text parsing is brittle.
- ✅ Test with the smallest/cheapest model first. Only switch to a larger model if the small one fails on real input.
- ⚠️ The token cost is the sum of the system instruction, all context memory, the request, and the response. Long context inflates cost on every run.
- ❌ Do not embed credentials or other secrets in the system instruction. They are sent to the AI provider on every call.
What to do next
- For multi-step reasoning with knowledge access, use AI Agent instead.
- To search a knowledge base without invoking a model, use Knowledge Query.
- For everything model-related: Models → Introduction.