skill-shortener

Run the skill-shortener workflow: it measures the per-section footprint, classifies each section as KEEP, CUT, or MOVE, presents a plan for approval, then rewrites the body with conditional pointers to new references/ files and verifies the result with automated checks.

Progressive disclosure keeps only universally needed content in the SKILL.md body and moves everything else into references/, scripts/, or assets/ behind pointers stating when to load them. This reduces the token cost paid on every activation.

The skill enforces two caps: 500 lines and 3000 words for the body, and both must be satisfied. The measure_skill.py script reports the current footprint and returns a WITHIN_CAP or OVER_CAP verdict against these limits.

Yes, Mode 2 (audit only) runs the measurement and classification phases, then presents the split plan and stops. It writes nothing outside the .skill-shortener workdir and skips the snapshot and git sync steps.

Do not use it for authoring new skills, fixing eval scores or frontmatter validation failures, or general prose editing. It also exits early when the body is already within both size caps, since splitting a fitting body only adds pointer overhead.

Every section is recorded in a manifest with its disposition, and verify_shorten.py checks that the manifest is exhaustive, all MOVE destinations exist and are pointed to with load conditions, no references are orphaned or chained, and the version was bumped. A manual read-back confirms content survived the move.