Skip to main content
Improve is the prompt improvement workflow in Adaline. It starts from production evidence, generates candidate prompt changes, tests them against evaluators and datasets, and stops for human review before the change is applied. Improve page showing pending review, in progress, and history cycles Use Improve when a prompt is already producing traces and you want a controlled loop from observed behavior to a reviewed prompt version. It is especially useful when a Behavior, Monitor chart, trace search, or evaluator result points to a repeated prompt-level issue.

​
What an Improve cycle does

An Improve cycle is a background run attached to one prompt in one project. The app shows the run as a cycle with five product stages:
StageWhat happens
BehaviorsAdaline reads clustered production behavior for the selected prompt and selected focus behaviors. If you did not choose a Behavior, the cycle can auto-pick high-priority failure patterns.
EvalsThe cycle inspects active evaluators linked to the prompt and any production-derived auto-evaluators available for the issue. Draft evaluators created by an in-flight Improve cycle are isolated until the cycle is approved.
DatasetsAdaline prepares train and validation cases from linked datasets, production traces, and generated edge cases when available. These cases are used to compare the baseline prompt against candidates.
PromptsThe optimizer generates multiple candidate prompt snapshots and blocks candidates that regress previously passing checks. Candidate diffs can include message changes, config settings, response schema changes, and tool-related changes.
ReviewThe reviewer sees the diagnosis, candidate switcher, prompt diff, real-traffic examples, regression report, cost, token, and latency summary, and audit packet.
The result is not applied silently. A reviewer chooses one of three outcomes:
  • Approve applies the selected candidate to the prompt and deploys it to the project’s primary deployment environment when one is configured.
  • Edit & approve applies the selected candidate and opens the prompt for editing without deploying.
  • Reject leaves the prompt unchanged and records the cycle in history.

​
Cycle states

The Improve page groups cycles by review status:
ViewWhat it shows
Pending reviewCompleted cycles with a candidate waiting for a decision. These are the highest-priority items for reviewers.
In progressQueued or running cycles. Open one to watch stage progress, live logs, cost, and generated evidence.
HistoryCycles that were approved, edited, rejected, failed, or canceled. Historical cycles keep the review evidence for audit and comparison.
Cycles can also be awaiting evidence when the prompt does not yet have enough evaluation cases or behavior evidence. In that state, Adaline waits for the evidence-generation pipeline to create usable eval coverage from production traces.

​
What Improve needs

Improve works best when the project has these ingredients:
IngredientWhy it matters
Production tracesGround the diagnosis in real requests and let Behaviors identify repeated patterns.
BehaviorsProvide semantic clusters and issue signals so the cycle can target high-impact patterns instead of isolated examples.
EvaluatorsDefine the checks that candidates must improve or preserve. Evaluators are the main guardrail against regressions.
DatasetsProvide stable examples for baseline and candidate comparison.
Deployment environmentsLet Approve publish the new prompt version to the primary environment. Without an environment, the candidate can still be applied and deployed later.
If a prompt has no clustered behavior evidence yet, collect traffic first and let the Behaviors pipeline run before starting an improvement cycle.

​
Review discipline

Treat Improve as a review workflow, not an autopilot. Before approving, confirm that:
  • The diagnosis matches the real product issue.
  • The selected candidate is the one you intend to ship.
  • Regressions are understood, especially evaluator drops and cost or latency increases.
  • The confirmation dialog names the deployment environment you expect.
  • You know how to roll back the environment if production behavior regresses.
  • The evidence that should become long-term regression coverage is captured in datasets.

Start an Improve cycle

Choose the prompt, focus area, target behaviors, and thoroughness.

Review candidates

Inspect the proposed change before creating a new prompt version.

Behaviors

Understand the behavior evidence Improve can target.

Evaluators

Define the criteria Improve uses to reject regressions.

Regression coverage

Promote production failures and Improve evidence into durable tests.