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

# Understand skill health

> Read SelfTune health states without confusing missing evidence with poor performance.

Skill health is a summary of the evidence SelfTune has observed. It is not a
permanent grade and it is not meaningful without enough representative usage.

## Ask your agent

```text theme={null}
Use SelfTune to explain the health of my installed skills. Separate observed
problems from no-data states, identify the evidence behind each conclusion, and
recommend at most one next action per skill. Do not modify any skill.
```

## Inspect health manually

```bash theme={null}
selftune sync
selftune status
```

Or open the local dashboard:

```bash theme={null}
selftune dashboard
```

| State             | Meaning                                      | Reasonable next action                               |
| ----------------- | -------------------------------------------- | ---------------------------------------------------- |
| Healthy           | Current evidence meets the accepted checks   | Keep observing                                       |
| Warning           | Evidence suggests drift or a bounded problem | Inspect the cited sessions or evals                  |
| Critical          | A current check failed materially            | Pause automatic changes and diagnose                 |
| Unknown / no data | Evidence is absent or insufficient           | Use the skill naturally or add a representative eval |

## Read the evidence, not only the label

Check the observation count, time window, skill version, harness, and kind of
evidence. A high score from a tiny or synthetic-only sample may be less useful
than a lower score from representative work. A no-data state is honest and
should not be converted into a failure.

## Success criteria

You should be able to answer:

* Which skill needs attention now?
* What evidence supports that conclusion?
* Is the issue activation, execution, output, or insufficient data?
* What is the smallest safe next action?

This is a read-only workflow. Sync only local evidence the user has approved.
For each skill, report the health state, evidence type, sample size or coverage
when available, time range, and uncertainty. Separate observation from
inference. Do not recommend an improvement when the correct state is no data.