Flow.Documentation

Building block: AI prompt

6 steps± 4 minVideoVerified 22 Oct 2026
  1. Step 1 · Open your development environment
  2. Step 2 · Create an AI prompt
  3. Step 3 · Write the prompt
  4. Step 4 · Test with a message
  5. Step 5 · Publish a version
  6. Step 6 · Give it to an agent

An AI prompt is a building block with one AI call following a fixed template: summarise, classify, pull data out of text. You fix the instruction and the template; whoever uses it only gives the input. That way every workflow and every agent gets the same, tested prompt.

Watch the whole guide as a video6 steps in one film
Step 1 / 6

Open your development environment

In the portal, under Environments, click Development, and there Open. Flow opens in a new tab and you are signed in straight away.

  1. Click Development
Open your development environment · Click Development
  1. Click Open ↗
Open your development environment · Click Open ↗
Step 2 / 6

Create an AI prompt

Click + next to General and choose AI prompt. Under Definition give it a name and a description, and an input field message: the text to summarise.

  1. Click Add to General
Create an AI prompt · Click Add to General
  1. Click AI prompt
Create an AI prompt · Click AI prompt
  1. In Name, type Summarise a customer message
  2. In Description, type Summarises a customer's message in one sentence.
Create an AI prompt · In Name, type Summarise a customer message; In Description, type Summarises a customer's message in one sentence.
  1. Click + Input field
Create an AI prompt · Click + Input field
  1. In Name, type message
Create an AI prompt · In Name, type message
Step 3 / 6

Write the prompt

Under Prompt, choose the AI provider, write the system instruction (how the model behaves) and the prompt itself. {{input.message}} fills in the input. Choose JSON object under Output when a next step needs fields. Save the draft.

  1. Click Prompt
Write the prompt · Click Prompt
  1. In AI provider, choose Scripted AI (test street)
  2. In System instruction, type You are a precise assistant for a bike shop. Answer in English.
  3. In Prompt, type Summarise this customer message in one sentence: {{input.message}}
Write the prompt · In AI provider, choose Scripted AI (test street); In System instruction, type You are a precise assistant for a bike shop. Answer in English.; In Prompt, type Summarise this customer message in one sentence: {{input.message}}
  1. Click Save draft
Write the prompt · Click Save draft
Step 4 / 6

Test with a message

Under Test, give a sample message and click Run test. In this guide a fixed test model answers, which returns the prompt; with a real model you get the summary here. The logs show which model answered and how long the answer was.

  1. Click Test
Test with a message · Click Test
  1. Click Code
Test with a message · Click Code
  1. In Input, type {"message": "My new e-bike stopped charging yesterday. The battery stays at 20 percent. Can I come by?"}
Test with a message · In Input, type {"message": "My new e-bike stopped charging yesterday. The battery stays at 20 percent. Can I come by?"}
  1. Click ▷ Run test
Test with a message · Click ▷ Run test
Test with a message · Passed, with the duration.
Test with a message · In the logs: the model and the number of characters.
Step 5 / 6

Publish a version

Publish a fixed version: that way the prompt never changes unnoticed under a workflow or agent.

  1. Click Publish version →
Publish a version · Click Publish version →
Publish a version · Version v1 is published and active.
Step 6 / 6

Give it to an agent

Open an agent and tick the prompt under Brain → Tools. An AI prompt only reads, so the agent may use it even when it may not change anything.

  1. Click Agents
Give it to an agent · Click Agents
  1. Click the agent ‘Customer service helper’
Give it to an agent · Click the agent ‘Customer service helper’
  1. Click Summarise a customer message
Give it to an agent · Click Summarise a customer message
  1. Click Save
Give it to an agent · Click Save
Give it to an agent · The building block is on as one of the agent's tools.

This guide was carried out step by step on 22 Oct 2026 by an automated test on Flow (version 3e563176).