Self-Improving

Watch the workflow: Read an authored suggestion and job history · Authored dummy data

Inspect an authored prompt suggestion in Tensile, compare its current and proposed skeletons, then expand the corresponding job history. No replay or model call ran. The displayed accepted gate, pass rates, safety and replay counts were written by hand; they are not measured improvements. No PR was opened, prompt applied, Test Heal run or candidate sent. This is a read-only layout walkthrough using dummy data.

Read the walkthrough transcript

• 00:00 — Review the authored prompt suggestion in Tensile. No replay or model ran.
• 00:05 — Inspect job history. Displayed pass rates are authored, not measured improvements.

Open Self-Improving to review two kinds of work. The healer proposes prompt fixes and tests them by replaying failing inputs. The miner proposes evaluator candidates from production traffic.

This page is available after your workspace has at least one healer job or one mined candidate. Without either, the page says Not enabled for this workspace. Ask your workspace administrator.

The healer only suggests. It does not change your prompts unless you choose to apply a suggestion.

Choose a viewCopy Anchor Link

The page has five tabs: Overview, Suggestions, Mined Evals, Job History and Configuration.

• Overview shows the number of jobs that are ok, failed and skipped, and a list of recent jobs. It says that no repository is connected and that the healer needs none.
• Job History lists every healer run. Open a run to read its root cause, the affected files, the reason for a failure or a skip, and the eval-gate evidence.
• Configuration says that per-workspace healer settings are not available yet and that platform defaults apply.

Review a suggestionCopy Anchor Link

1. Open Suggestions.
2. Read the card. It shows the rubric, the replay result as a before and after pass rate, the number of replays, and whether safety held.
3. Compare the current and the proposed prompt skeleton.
4. Select Copy proposed prompt to copy the accepted skeleton.
5. Select Apply to registry to save it as a new prompt version. The card then shows the agent and version.

A suggestion that the eval gate rejected stays in a separate list. It has no apply control.

Decide on a mined candidateCopy Anchor Link

1. Open Mined Evals.
2. Read the title, the rubric, the reason and the number of supporting cases.
3. Select Edit to change the title or rubric, then Save.
4. Select Reject to dismiss the candidate.
5. Select Send to evaluator packs to start it as a discovered pack for validation.

A candidate and a discovered pack are not used for scoring. After you send a candidate, the card shows the start of the pack ID.

Empty statesCopy Anchor Link

• No heal jobs yet appears when the healer has not queued a job. The healer queues one when a rubric pass rate drops on live traffic.
• No suggestions yet appears until the healer verifies a fix.
• No mined candidates yet appears until the miner finds recurring disagreement or anomalies.

Next stepsCopy Anchor Link

Review the checks that you sent in Evaluator packs.