Prompt

Send instructions and content to an LLM and use the response.

Prompt is the workflow's direct channel to an LLM. It sends instructions, a query, and optionally documents to a model and returns the response — the way you bring generation and reasoning into an automated process.

Reference layer pending. This page documents the skill's purpose and configuration as used in the builder. Its formal configuration schema, output schema, and error codes are not yet documented — see the Skill Library overview.

Worked example: summarizing customer feedback

A workflow collects feedback from survey responses and support tickets. The team needs a concise summary of each — key themes, overall sentiment — rather than a pile of raw text.

The problem. Reading and summarizing large volumes of feedback by hand is slow and inconsistent between whoever does it.

The approach. Send each piece of feedback to an LLM and ask for a summary.

1. Choose the model. Pick from the Model dropdown — Gemini-2.5-Flash, gpt-4, and a continuously growing list.

Selecting the model

2. Write the instruction. Instruction to LLM (formerly "System Prompt") sets the model's role for this task:

You are a helpful assistant specialized in summarizing customer feedback. Your task is to provide a concise, neutral summary of the user's input, highlighting key points and overall sentiment.

Giving the model its role and task

3. Set the query. Point User's Query at the text to summarize — $input.customerFeedbackText.

4. Attach documents (optional). When the response needs extra context — a product manual, a policy — attach them by Document ID(s) (comma-separated) or Base 64 Content.

Advanced configuration

Defaults are usually right. Change these when you've measured a reason to:

Setting	What it does
Temperature	Controls creativity/randomness in the response. Lower is more predictable — better for factual work.
Number of Conversation Turns	How many previous messages the model should consider. Default 3.
Response Format	Plain text by default. Set to JSON when a later step parses the output.

Output

• content — the model's generated response.

• systemPrompt — the Instruction to LLM that was sent, returned for traceability.

• error — descriptive message; null on success. Particularly useful when responseFormat is JSON and the model failed to produce valid JSON.

• statusCode — 200 success · 400 bad request · 500 service-side failure.

Setting Response Format to JSON asks the model for valid JSON; it does not guarantee it. Check error before a downstream step tries to parse content, or the failure surfaces somewhere less obvious.

---

To add this skill to an agent, see Adding a Skill to the Agent. To ground responses in your own documents rather than the model's training, see RAG Query.