> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trulayer.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Prompt Improvements

> Closed-loop prompt improvement: review model-proposed prompt diffs, ship them after an A/B replay passes, and roll back automatically when a shipped prompt regresses.

The **Prompt Improvements** page closes the loop on the failures TruLayer observes in production. The platform clusters production failures, asks an LLM to synthesize a candidate prompt diff, replays a held-out trace set against both the live prompt and the candidate, and proposes the diff for review.

## State machine

A **prompt deployment** moves through a fixed state machine. Every transition is recorded; you can see the current state on every row of the list.

| State         | What it means                                                                                          |
| ------------- | ------------------------------------------------------------------------------------------------------ |
| `proposed`    | The cluster→diff worker synthesized a candidate prompt. Awaiting A/B replay.                           |
| `ab_running`  | The A/B harness is replaying the held-out trace set.                                                   |
| `ab_passed`   | A/B replay finished and the candidate is recommended for ship.                                         |
| `ab_failed`   | A/B replay finished and the candidate did not improve the metric set.                                  |
| `shipped`     | An owner approved the candidate and it is now live for the project.                                    |
| `monitoring`  | The shipped prompt is being observed against the post-ship rolling window.                             |
| `regressed`   | The rolling window crossed the regression threshold. A regression banner appears across the dashboard. |
| `rolled_back` | An owner (or the auto-rollback monitor) rolled the shipped prompt back to its predecessor.             |
| `rejected`    | An owner rejected the candidate before it was shipped.                                                 |

## Reviewing a candidate

Click any row to open the detail page.

The detail page shows:

* **Candidate diff** — a unified-diff view of the proposed prompt against the currently shipped prompt. Added lines are highlighted green; removed lines red.
* **A/B replay report** — per-metric baseline, candidate, and delta. A passing report (green badge) is the default precondition for shipping.
* **Decision controls** — owner-only buttons to approve, reject, or roll back.
* **Post-ship monitoring** — the rolling regression metric over time. A dashed reference line marks the A/B candidate baseline so divergence is visible at a glance.

## Approving and rolling back

Approve a candidate by clicking **Approve & ship**. The prompt becomes live immediately and enters the `monitoring` state. Until you approve, no production traffic is affected — the candidate is sandboxed in the A/B replay only.

Reject a candidate or roll back a shipped prompt by clicking **Reject** or **Roll back**. Both require a free-text reason that is recorded on the deployment.

Only users with the `owner` role can approve, reject, or roll back. See [Team settings](./settings-team) for roles.

## Regression alerts

When a shipped prompt's rolling regression metric crosses the configured threshold, the deployment moves to `regressed` and a banner appears across every dashboard page. Click the banner to jump to the regressed deployment and decide whether to roll back manually or let the auto-rollback monitor act.

## Plan availability

Closed-loop prompt improvement is part of the **Control plane** feature set and is gated on the **Team** plan or above (same gate as Control, Policies, Alert Rules, and Model Routing). Starter and Pro projects see an upgrade CTA in place of the page.