Claude Code normally sends the description of every installed skill with every request. With the picker on, that list is replaced by a one-line note, and Jev attaches a shortlist of the few skills each new prompt actually needs.
In global scope, Claude may read the shortlisted skills' folders without asking. The roster is cached in .optim-jev/jev_skills/index.json; /jev-skills reindex rebuilds it.
- Skip? No shortlist for slash commands, prompts under 12 characters, or follow-ups like "yes", "ok", "continue".
- Trim: over 40 skills, a keyword match against the prompt keeps the 40 closest.
- Batches: skills are split into groups of 20. Jev picks the best fit in each group, with "none" always an option.
- Final round: with more than one group, the top 2 of each (at 15% or more) are scored together once more.
- Shortlist: up to 3 skills that beat "none" and reach 30%.
| Result | What Claude receives |
|---|---|
| shortlist | the skills with their descriptions and percentages, and a note to load one if it fits |
| no match | nothing extra |
| Jev offline | the names of every skill in scope |
Jev sees the prompt and your 3 prompts before it (your own text only), so follow-ups are judged in context. Each request to Jev is capped (jev_skills_max_input_tokens, default 3000): earlier prompts are dropped oldest first, and skills are split into smaller batches so each batch fits.
- The keyword trim can drop a relevant skill whose description shares no words with the prompt.
- Shortlist notes are new text on each prompt, so they're written to the prompt cache rather than read from it.