← All Skills
🧪
Skill Conductor Evaluation Lifecycle
A full lifecycle method for creating, improving, validating, reviewing, and packaging Agent Skills, with architecture selection, no-skill baselines, binary evidence evaluation, blind comparison, held-out gates, and bounded self-update.
Key Features
- Six skill-lifecycle modes
- Baseline and trigger pressure tests
- BinEval binary evidence scoring
- Held-out gated self-update
Installation
Run npx skills add smixs/skill-conductor. Claude Code can instead add the smixs/skill-conductor marketplace and install skill-conductor@smixs.
How to Use
Ask to create, repair triggering, evaluate, review, optimize, or package a skill. It identifies the mode and failure class, then applies baselines, structural checks, with/without-skill comparisons, and evidence-gated improvements.
Requirements
Script-backed modes require uv. Separate eval and improvement runs also need a logged-in Claude CLI or ANTHROPIC_API_KEY; optimization requires the Claude CLI. Test prompts and outputs sent to an external model may leave the machine, so keep secrets and sensitive project data out of evals. Without model access, users can run tests themselves and return the outputs.
Related Skills