every turn loads the skill index into the prompt, so big skill bodies are pure tax. the plan was a library wide slimming pass: one full backup for rollback, then a fan-out of one worker per skill, each with the same job and a strong rule, no folder gets deleted, every line of detail lands somewhere.
the pattern in each: keep the triggers, the workflow steps, commands, paths, and the sharp pitfalls in the top of the file, and move the long tables, api catalogs, historical fixes, per-device deep dives and troubleshooting trees into reference files under the same skill. a couple of the biggest files go from over sixty thousand characters down to a slim core without losing a word, because the content moved, nothing was destroyed.
then the duplicate pass merged a companion file into its parent and collapsed four overlapping skill files into the one we actually load, keeping the endpoint, the cache rules, and the platform quirks. checksums of the reference files matched the originals, nothing outside the named folders changed, and the totals dropped hard. a smaller skill body is a smaller cost per load, and a shorter index is a shorter header on literally every conversation that starts. the most boring afternoon of this project and one of the most durable.