Skill Compressor
Longer agent instructions can consume more context without producing better behaviour. Shorter instructions can also delete the line that made a skill reliable.
Skill Compressor makes a skill cheaper to load without breaking when it activates or how it behaves. It freezes a baseline, proposes a smaller candidate and accepts it only when routing and runtime tests show that the important behaviour survived.
Use it when
- A skill has grown expensive to load.
- Repeated explanations belong in progressive disclosure.
- You need to distinguish useful instruction from accumulated prose.
- A shorter version must be supported by evaluation, not taste.
Example request
Use skill-compressor to reduce this skill's expected token load without changing when it activates or what it does.What to expect
The result includes a measured before and after comparison. A shorter file does not win if it loses required behaviour.
Install and source
npx skills@latest add simonasrazm/skills --skill skill-compressorRead the Skill Compressor source or the skills.sh listing.