skill-evolver
skill-evolver
Plan improvements for an existing larch skill in a single research-and-file-issues pass. Take a mandatory <skill-name> (must already exist under skills/<name>/ or .claude/skills/<name>/ in the current plugin repo), invoke /research via the Skill tool against repo-local sibling skills + reputable external sources (Anthropic, OpenAI, DeepMind, ≥500-star OSS), and — if the research lane surfaces ≥1 actionable improvement with citations — invoke /umbrella via the Skill tool to file the resulting GitHub issue(s). /research runs the fixed 4 research + 3 validation lane shape internally. /umbrella runs its own one-shot vs multi-piece classifier on the distilled task description.
The skill itself does NOT modify the target skill's files. Implementation of each improvement happens later via /fix-issue. This skill is research-and-file-issues only.
Example: /skill-evolver design or /skill-evolver review.
Before editing, read
${CLAUDE_PLUGIN_ROOT}/skills/shared/skill-design-principles.md(full file). Section III mechanical rules A/B/C override general writing-style guidance on conflict.
Prerequisites
/skill-evolver delegates to two sibling skills via the Skill tool. Both must be present in the loaded plugin/session:
/research— ships with this plugin under${CLAUDE_PLUGIN_ROOT}/skills/research/(always available when the larch plugin is loaded)./umbrella— ships with this plugin under${CLAUDE_PLUGIN_ROOT}/skills/umbrella/(always available when the larch plugin is loaded).
Anti-halt continuation reminder. After every child Skill tool call (/research, /umbrella) returns, IMMEDIATELY continue with this skill's NEXT numbered step — do NOT end the turn on the child's cleanup output, and do NOT write a summary, handoff, or "returning to parent" message — those are halts in disguise. The rule is strictly subordinate to any explicit non-sequential control-flow directive in THIS file (exit cleanly, skip to Step N). → shared/subskill-invocation.md#anti-halt