Designing an Experiment with the AI Assistant

Setting up a new experiment usually means translating what you know about your problem into parameters, objectives, and constraints. With the assistant, you don't need to arrive with that translation done. You start with a single sentence and the assistant guides you through what needs to be defined, one step at a time.

Tip: If you already have your experiment written up somewhere — a protocol, a report, a lab notebook entry — you can paste it directly into the chat instead of starting from a sentence. See Starting from an Existing Experiment Description.

Open the assistant from the SDLabs front page, or from the assistant icon in the top right of a draft experiment. Then describe your goal in plain language:

"I would like to optimize my chocolate cake recipe."

That's enough to begin. The assistant takes it from there, asking for what it needs in the order it needs it. As it configures each piece, the draft experiment updates in the interface in real time, right next to the conversation — so you can watch the design take shape as you talk.

The assistant asks what you can vary and over what range:

"The oven temperature between 150 and 220 °C, the baking time up to 60 minutes, and a choice of four types of chocolate."

From a single answer like this, it sets up the numerical parameters with their ranges and the categorical one with its allowed options. If something is missing — a unit, a lower bound — it asks rather than assuming.

If your categorical options have known properties (descriptors) that could help the optimizer, the assistant can add them — using the values you provide.

Next, the assistant asks what you will measure and whether each result should be maximized, minimized, or hit a target.

If you have more than one goal — say, the moistest cake at the lowest cost per batch — it helps you decide how to balance them:

You don't need to know these strategies by name; describe what matters to you and the assistant proposes the right one.

You state your constraints in plain words:

"Sugar and cocoa together can't exceed 40% of the batter, and don't pair dark chocolate with the highest oven temperatures."

The assistant maps each statement to the right type of constraint. No formulas needed on your side.

Typical ranges, relationships you've observed, combinations that look promising — this kind of knowledge helps the optimizer focus its search. The assistant drafts it into a short expert-context summary.

It can also draw on its own scientific knowledge to help you write it well — suggesting relationships worth capturing or sharpening how a known effect is phrased — but it always checks with you, and it never asserts something about your system that you haven't confirmed.

Finally, you choose how many candidate conditions you want suggested per round (the batch size), and review the configured experiment side by side with the conversation.

Launching is always your click. The assistant prepares the draft, but it never starts or publishes an experiment on its own. When you're satisfied with the design, you take it live yourself.

Ambiguous or missing information triggers a clarifying question, never a silent guess. If you give a range without a unit, or an objective without a direction, expect the assistant to ask — that's by design.