Improve a skill
Improve an existing skill from its own history: the redirects and errors that happened while it was loaded. The new version is staged beside the old one, and the counters decide.
Improve
Section titled “Improve”rematch improve <skill> # e.g. forge-portfolio-weightingrematch improve <skill> --split-only # just the mechanical split, no modelIn omp: /rematch improve <skill>.
What it does
Section titled “What it does”The skill is registered as a provisional entry in place (its file is never moved), and the sessions that read it are audited. Then one of three things happens:
| Situation | Result |
|---|---|
| fewer than 5 traced sessions used it | refuses with the count and offers the split |
| over 8 KB | split mechanically, no model: router SKILL.md + units/ by heading + facts/, staged beside the original |
| otherwise | one editor call proposes unit-level edits from the clustered redirects and tool-error clusters; same lint; staged beside |
- Imported sessions count too: their excerpt is rebuilt from omp’s session files.
- The split also moves long tables and rejected-hypothesis lists into a ledger-notes unit.
- Until the counters decide, omp shows the new version under the bare name and the original under its namespaced name.
Before a new version replaces a kept one, its originating task is replayed once if it has a verifier: replacing a kept skill.
Import compound-forge skills
Section titled “Import compound-forge skills”rematch import --forge <path>rematch import --forge <path> --migrate- Registers compound-forge’s skills as entries, with predicates from their trigger text.
- Registers its
learnings/*.mdas lessons with their evidence hash. --migratecopies the skill trees intostaging/forge/(scripts non-executable), so the forge plugin and checkout can be removed. The forge’s own meta-skill stays behind.
This is how compound-forge was retired on both of the author’s machines.