AI Prompts

Use this section to copy proven prompt patterns into Torii Workflows using AI (GPT) Actions. Each prompt guide is designed to be:

• Reusable across customers (generic inputs + deterministic outputs)
• Workflow-friendly (structured responses you can map to tokens and branch on)
• Safe by default (clear constraints, predictable formatting)

Note: These guides focus on AI (GPT) Actions (Torii’s built-in GPT action). They’re written for Torii customers, partners, and internal builders who want repeatable automation patterns. The documentation can be found here.

How to use these prompts in Torii

At a high level, each prompt guide follows the same implementation flow:

1. Pick a Workflow trigger
   Example: New application discovered, Application meets criteria, User meets criteria, etc.

2. (Optional) Normalize data with Transform Data
   Use this action to shape raw trigger/app/user data into clean inputs (tokens) for the AI prompt.

3. Add an AI (GPT) Action

   • Paste the prompt from the guide
   • Define Response mapping (required) to capture the model output into workflow tokens
   • Use those tokens in subsequent actions (e.g., update application fields, route approvals, post to Slack)

Response mapping is mandatory—without it, the output can’t be used downstream.

Prerequisites and environment setup

AI Settings and OpenAI keys

Torii runs AI (GPT) Actions using OpenAI. You can use Torii’s shared access, or configure your own OpenAI API key for more predictable usage and higher limits based on your OpenAI account.

➡️ Manage AI Settings (Support):
https://support.toriihq.com/hc/en-us/articles/38227217845531-Managing-AI-Settings

Benefits of providing your own key include:

• Higher rate limits (avoid limitations from Torii’s shared key)
• Ability to use options like Search the web (when applicable to the action)

Note: Torii stores the provided OpenAI API key securely and it isn’t accessible to users.

Prompt design principles

1) Make outputs machine-usable

Prefer outputs that are easily used in downstream actions:

• Enumerations (e.g., SSO | No SSO | Unknown)
• Booleans (true/false)
• Simple text strings
• Small JSON objects with fixed keys

2) Constrain the model

Every prompt in this section should:

• Define allowed outputs (and forbid anything else)
• Specify what to do when uncertain (e.g., return Unknown)
• Avoid open-ended prose unless the workflow truly needs it

3) Treat prompts like workflow contracts

A prompt is only “good” if it stays stable over time:

• Keep them as short and explicit as possible.
• Prefer deterministic formatting.
• Include examples when helpful.

When to enable “Search the web”

AI (GPT) Actions can optionally “Search the web” for real-time context when you have configured your own API key. Enable it only when the task truly depends on up-to-date public information (e.g., recent policy changes, current company details). Otherwise, keep it off for faster and cheaper runs.