skill-router-builder: group 380 skills into 36 routers, measured on two public collections

[gorkicakir]

If you install many skills, every conversation starts with the name and description of each one. Claude Code also caps that listing at about 1% of the context window and silently drops the descriptions of the skills you use least, so with 100+ skills most of them become invisible to the model.

skill-router-builder reads a skills folder, clusters related skills by file links, mentions and text similarity, and writes "router" skills whose maps are generated from the members' own descriptions. Nothing is written by hand, and the originals are untouched.

Measured on two public MIT collections, tool unchanged (default settings):

Collection	Top-level skills	Listing size (tiktoken)	Correct skill picked
K-Dense-AI scientific skills (177)	177 -> 112	14,867 -> 9,573 tokens (-36%)	35/35 flat, 35/35 routed
Alireza Rezvani business/engineering skills (380)	380 -> 36	19,566 -> 4,612 tokens (-76%)	35/35 flat, 35/35 routed
My own collection (78)	78 -> 6	~14,700 -> ~1,200 tokens (-92%, estimated)	69/70 both

No false triggers in either public collection. The scientific set groups less because 102 of its skills are unrelated single-library wrappers; the tool does not force them together. Full tables and the 35 requests per collection are in tests/RESULTS.md.

Install

pip install "skill-router-builder[recommended] @ git+https://github.com/gorkicakir/skill-router-builder"
skill-router --src ~/path/to/skills --out out --dry-run

Python 3.8+, stdlib only; networkx is recommended (the fallback clustering groups noticeably less). MIT.

https://github.com/gorkicakir/skill-router-builder

Happy to hear about collections where the grouping comes out wrong.