Skills Commands¶
superqode skills manages and optimizes project skills. A skill is a markdown artifact that teaches the harness a repeatable workflow, review style, tool sequence, or project convention.
superqode skills COMMAND [ARGS]...
optimize¶
Optimize a markdown skill with GEPA. The live skill is never overwritten; SuperQode writes staged artifacts and a review report.
superqode skills optimize review \
--engine omni \
--harness harness.yaml \
--tasks eval-tasks.yaml \
--live
--engine gepa keeps the stable reflective optimizer. The engine-pluggable API also supports autoresearch, gepa-meta-harness, and omni. Omni runs GEPA, AutoResearch, and GEPA's meta-harness engine in parallel, selects the best result, and gives that candidate a fresh continuation run.
Important options:
| Option | Purpose |
|---|---|
--engine TEXT | gepa, autoresearch, gepa-meta-harness, or omni |
--continuation-engine TEXT | Fresh engine used after Omni exploration |
--harness PATH | Harness used to evaluate candidates |
--tasks PATH | Eval task file |
--output PATH | Directory for staged artifacts |
--provider TEXT | Provider for eval runs |
--model TEXT | Model for eval runs |
--runtime TEXT | Runtime or backend override |
--sandbox TEXT | Sandbox profile, default local |
--reflection-lm TEXT | Model used by GEPA for reflection |
--max-metric-calls INTEGER | Evaluation budget |
--explore-max-evals INTEGER | Per-engine exploration budget for Omni |
--max-token-cost FLOAT | Optimizer-model USD cap, partitioned across Omni stages |
--optimizer-model TEXT | Claude model for AutoResearch and meta-harness |
--agent-sandbox / --no-agent-sandbox | Sandbox agent-engine subprocesses |
--max-edits INTEGER | Bounded edit limit |
--live | Execute eval tasks against the configured model |
--json | Emit JSON |
Tasks marked split: held-out are sealed from optimization and used for the final non-regression gate.
For workflow guidance, see Skill Optimization.