skill-improver - Optimize skills through validated experiments
Runs a skill against binary evaluations, diagnoses failures, tests controlled edits, and keeps only validated improvements.
Tags
Updated: 2026-10-01Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Audit target skills
- Gather evaluation setup
- Establish performance baselines
- Score outputs with binary evals
- Diagnose failures from full traces
- Select Pareto frontier parents
- Propose controlled skill edits
- Validate edits with PACE
- Seal optimized skill versions
Inputs
- Target skill path
- Test inputs
- Binary evaluation criteria
- Model configuration
- Runs per experiment
- Budget cap
- Golden cases
- Output version name
- Existing checkpoint
Outputs
- Improved skill copy
- Baseline skill copy
- Results TSV log
- Mutation changelog
- Live HTML dashboard
- Execution traces
- Checkpoint files
- Final test score
Requirements
- Target skill files
- Skill-audit command
- File read and write access
- Configured optimizer model
- Configured target model
- Python runtime
- Browser access