Skip to main content
Trigger an Improve cycle when you want Adaline to propose a prompt change from production evidence. A cycle runs in the background and moves to Pending review only after it has produced reviewable evidence. The best cycles start with a clear pattern: a Behavior, trace set, evaluator failure, or dataset gap. Improve wizard with prompt, focus, behavior targeting, thoroughness, and notification controls
1

Open Improve

Open a project, then select Improve.
2

Choose the prompt

Click Start improvement and select the prompt Adaline should improve.
3

Add a focus

Describe the change you want. Name the failure mode, audience, policy, output format, tool behavior, or runtime tradeoff.
4

Target behaviors

Select one or more Behaviors when you want the cycle to focus on known production patterns.
5

Pick thoroughness

Choose Quick, Standard, or Thorough based on risk and urgency.
6

Start the cycle

Add reviewers to notify, then click Start cycle.

​
Choose thoroughness

Actual runtime depends on prompt size, trace volume, dataset size, evaluator cost, provider latency, and candidate count.
Unless your billing plan includes cycle runs, Improve cycles are charged per run. The free plan does not include cycle runs. Typical cycle costs are about USD 2 - USD 5 per run, depending on the run configuration and usage.

​
Readiness checks

The run button may be disabled when Adaline does not have enough evidence. Common causes:
  • The prompt has not produced production traces.
  • Behaviors are not yet clustered for the prompt.
  • No useful evaluation or dataset coverage exists for the target issue. This coverage is optional, but weak coverage makes the review decision less confident.
  • The selected Behavior does not map cleanly to the selected prompt.
Fix this by sending traffic, improving instrumentation metadata, adding representative cases, or choosing a narrower Behavior.

​
After starting

The cycle appears under In progress while Adaline analyzes Behaviors, inspects evaluators, prepares datasets, explores candidates, and builds the review packet. You can leave the page while the run continues. Use the cycle list by status: If the cycle fails, open the failed run and read the failure reason. Do not treat a failed cycle as a product decision; it usually means the run lacked evidence, scoring coverage, or a valid prompt diff.

Review a Cycle

Inspect the completed evidence packet and make a release decision.

Synthetic Datasets

Improve validation coverage before rerunning weak cycles.