AI Prompt

Send one prompt to a model, without a saved agent — text, a fixed JSON shape, or your own schema.

The AI Prompt node (ai_prompt) sends one prompt to a model you pick and passes the answer on. Use it for single, self-contained steps — summarise, translate, extract fields, read a scanned document — where a full AI Agent with tools and memory would be more than you need.

Configuration

Field	Type	Default	Description
`llm_config_id`	string	—	The LLM provider connection to use, from Settings → LLM Configurations.
`model_id`	string	—	The model on that connection.
`instructions`	string	`''`	The system prompt. May contain `{{ }}` references. Studio keeps a version history of it; the active version is what runs.
`inputExpression`	string	`''`	The user message, with `{{ }}` references. Empty forwards what reached the node, the same way the AI Agent node does.
`attachmentExpression`	string	`''`	One uploaded file for the model, for example `{{ trigger.body.scan }}` from a Form file field or `{{ upstream.download.output.file }}` from an FTP download. An image or PDF is sent as the file itself, and the model must be marked vision-capable. Any other file (Word, Excel, CSV, text) is sent as its text, after the input, and works with any model.
`temperature`	number 0–2	`0.7`	Sampling temperature. Lower is more repeatable.
`maxTokens`	integer 1–128000	`4096`	The answer's token limit.
`outputFormat`	`text` \| `structured` \| `custom`	`text`	Text passes the answer through. Structured asks for, and checks, one of three fixed JSON shapes. Custom asks for, and checks, your own JSON Schema.
`structuredSchemaId`	`classification` \| `sentiment` \| `extraction`	—	The fixed shape for Structured: `classification` is `{ category, confidence, summary }`, `sentiment` is `{ sentiment, confidence }`, `extraction` is `{ fields: { … } }`.
`customJsonSchema`	string	—	A JSON Schema for Custom. It may itself be a `{{ }}` reference, so a caller can supply the schema per run.

To tell the model what day it is, put {{ run.startedAt }} in the instructions or the input.

Output

Field	Type	Description
`completionText`	string	The model's answer as text.
`structuredOutput`	unknown	The parsed, validated JSON, for Structured and Custom. Read fields as `{{ upstream.<id>.output.structuredOutput.category }}`.
`modelUsed`	string	The model that answered.
`tokenUsage.prompt`	number	Prompt tokens.
`tokenUsage.completion`	number	Completion tokens.
`latencyMs`	number	How long the call took.

The output also carries promptSent, the full prompt sent — the system and user halves, labelled — which the node's editor shows for debugging, and estimatedCostUsd, the call's estimated cost (absent when the model has no price).

Limits and errors

The node fails, naming the reason, when:

no LLM configuration or model is selected, or the configuration belongs to another organization;
the output format is Structured with no shape picked, or the answer does not match the shape (or your Custom schema) — the node never passes on an unchecked blob;
attachmentExpression points at an image or PDF and the model is not vision-capable, does not resolve to exactly one uploaded file, or the upload has expired;
the attachment is not an image or PDF and no text could be read from it (convert it to PDF, or run the step that produced it again).

A code fence around a JSON answer is removed before it is checked.

Example

OCR a scanned invoice from a form, then map it to fields:

workflows/read-invoice.yaml

name: Read invoice
description: Reads an uploaded invoice scan into fields.
graph:
  nodes:
    - id: upload
      type: trigger_form
      position: { x: 0, y: 0 }
      data:
        authMode: none
        fields:
          - {
              name: scan,
              label: Invoice scan,
              type: file,
              required: true,
              acceptedExtensions: [pdf, png, jpg],
            }
    - id: read
      type: ai_prompt
      position: { x: 280, y: 0 }
      data:
        llm_slot: vision-model
        model_id: gpt-4o
        instructions: Extract the invoice number, date, supplier and total.
        inputExpression: 'Today is {{ run.startedAt }}.'
        attachmentExpression: '{{ trigger.body.scan }}'
        temperature: 0
        maxTokens: 1024
        outputFormat: structured
        structuredSchemaId: extraction
  edges:
    - { id: upload-to-read, source: upload, target: read }

In a Package

The LLM connection does not travel: it becomes an llm Slot the installer binds to one of their own LLM configurations.

Travels: model_id, instructions, inputExpression, attachmentExpression, temperature, maxTokens, outputFormat, structuredSchemaId, customJsonSchema
Converted: llm_config_id → llm_slot (an llm Slot id)
Never carried: none

elizon.yaml

slots:
  - id: vision-model
    type: llm
    required: true