Evidence-based Skill optimization for coding agents.
skill-optimizer helps improve agent Skills using real usage logs. It looks at how a Skill is actually triggered and used, then helps produce an optimization report and a candidate Skill version.
Agent Skills tend to grow over time:
- old rules stay around after they stop mattering
- examples and FAQ sections become duplicated
- trigger descriptions get too broad
- low-frequency details stay in the always-loaded context
- repeated workflows remain as instructions instead of scripts
Blindly shortening a Skill is risky. A smaller Skill can make tasks fail.
Given a target Skill and historical Claude or Codex logs, skill-optimizer helps identify:
- whether the Skill is loaded for matching tasks
- whether loaded sessions show useful behavior
- which sections have deterministic evidence
- which content may be moved, rewritten, or removed
- which repeated workflows may become scripts
- how Skill metrics change across versions
This is not a prompt compression tool.
skill-optimizer separates:
- static Skill size reduction
- loaded-context reduction
- observed task-level impact after adoption
The goal is:
lower token cost per successful skill-related task
while avoiding regressions in task success.
Send this to your agent:
Install https://github.com/huang-hf/skill-optimizer as an agent Skill.
Send this to your agent:
Optimize my <skill-name> skill with skill-optimizer. Create a report and a candidate version.
Example:
Optimize my-test-skill with skill-optimizer. Create a report and a candidate version.
- Claude Code
- Codex
Prototype / MVP.
No third-party Python dependencies are required.