> ## 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.

# Improvements and proposals

> Understand how SelfTune turns a bounded problem into a reviewed, verifiable skill change.

An improvement is a measured change to a skill. A proposal is the reviewable
record that connects the problem, evidence, candidate diff, validation result,
and recovery path.

## The safe loop

1. Identify a specific problem supported by evidence.
2. Select the smallest relevant skill surface.
3. Generate or author a candidate without changing the accepted version.
4. Compare both versions under equivalent conditions.
5. Ask a human to approve consequential writes or publication.
6. Observe the accepted version and recover if it regresses.

Terms such as evolve, search, replay, and orchestrate describe implementation
paths inside this loop. You do not need to choose among them for ordinary use;
state the desired outcome and let the version-matched agent skill route it.

See [Improve a skill](/guides/improve-a-skill) to run the workflow.

Keep accepted and candidate revisions distinct. Record the evidence,
mutation surface, diff, evaluation conditions, result, approval, and rollback
path. Prefer no change when the comparison is ambiguous. Never publish or
replace the accepted version merely because a candidate was generated.