Thunk development typically follows three stages: build, iterate, and run in production. Now that you've built your thunk, the next step is refining it—testing with representative data to ensure your AI-powered workflow delivers the intended results before moving to production.
A core promise of the Thunk.AI platform is consistency and reliability in how your AI handles repeated tasks. This article goes into details of how the platform ensures reliable AI behavior. In this tutorial, we will walk you through practical steps to fine-tune your thunk, giving you control to steer your AI toward reliable, consistent, high-quality outcomes.
Recommended Settings for Iteration Mode
To maintain full control while refining your thunk, we recommend adjusting these AI Governance settings in the your thunk:
Enable Draft Mode – Prevents your AI from automatically processing new rows, allowing you to process data at your own pace and make refinements as needed.
Disable Auto-Complete – Ensures your AI doesn’t finish steps automatically, giving you the opportunity to assess and adjust its outputs.
Control & Steer Your AI with Precise Step Instructions
Reliable AI execution starts with a strong planning phase, where intent is captured through step-by-step instructions for both the overall workflow and individual tasks. The choices made during planning directly impact AI reliability, reinforced by Thunk.AI’s design to ensure precision and consistency.
Every request or data processed in a thunk follows the workflow, guided by Step Instructions set during planning. Refining these instructions allows you to steer your AI toward the desired outcomes.
Step Instructions: Your AI’s Primary Control Mechanism during Execution
On the Thunk.AI platform, Step Instructions go beyond simple prompts—they define how AI operates at each step. They include:
Instructions – Define what the AI should do and how to approach the task, including any user responsibilities.
Examples – Provide good and bad output examples to clarify expectations.
Property Bindings
Inputs – Specify the data properties or context the AI will use.
Outputs – Define the expected results of the step.
Tools – List the modules and tools available for the AI to use during execution.
Each of these elements is fully customizable, giving you fine-grained control over your AI’s behavior. Let's explore some examples.
Example: Refining Step Instructions for Predictable, High-Accuracy Results
We will use the same sample thunk for "Processing incoming purchase orders" as a running example in this tutorial.
Here are examples of refining your thunk's Step Instructions to ensure your AI processes purchase orders reliably and accurately.
1. Adjust Step Instructions for Clarity
Make instructions clear and specific to guide your AI effectively.
Example:
"Extract details from the purchase order." → Too vague.
"Extract product names, quantities, unit prices, total amount, and payment terms." → Clear and actionable.
2. Ensure Property Bindings Are Correct
In Data Bindings, your AI uses predefined inputs and outputs to process each step. Any property made available to your AI in a step will be populated accordingly, ensuring structured execution.
3. Refine Property Descriptions for Precision
Detailed property descriptions help your AI extract and populate the right information.
Example: If extracting the total order amount:
Type → Number
Description → "Total amount due, typically calculated as the sum of all product line items."
This ensures your AI extracts a valid number and correctly identifies the total order value from the order.
Example: Provide file context in input file property descriptions to clarify structure and intent. "This is an incoming purchase order from a client. Product details, quantities, and payment terms are listed in the table on page 2."
These refinements equip your AI with the context you’d typically use, improving accuracy, reliability, and workflow consistency.
How the platform uses Step Instructions to steer the LLM's work
When your Thunk AI agent does work on a step, it starts with "micro-planning" the task using the Instructions. The dynamic micro-plan is constrained by the available tools, by the specified input and output data bindings and by the provided examples. By explicitly requiring such articulation of the micro-plan, the Thunk.AI platform steers subsequent stages of the iteration in a consistent direction. Since the output bindings are schematized and structured, this further imposes constraints on the output of the LLM, increasing alignment of the LLM's responses with the desired outcomes. You can observe this dynamic micro-plan creation and execution in the AI log during or after work is done.
After you refine any aspect of the Instructions, examples, tools used and data bindings as shown above, you can re-run your AI on the Step using the circle icon.
Pro Tip: In iteration mode, start with a small set of data in your thunk. Once you're confident in your AI’s performance, scale up to process more data.
Process Data in Bulk: When you're ready to process data in bulk, use the rocketship icon in your Data Table to kick off multiple draft data entries at once. We recommend using this mode only after refining your step instructions and ensuring your AI delivers consistent results, helping you stay in control of usage and costs. You can find the rocket ship icon in the lower bar when you have Draft data entries in your Data.