jev-skill-suggestion

What it does

Claude Code picks a skill from a one-line listing of every skill it has. With a lot of skills, that listing gets truncated and the pick gets worse: it loads the wrong skill, or loads one when nothing fits.

This mod scores every skill against your prompt while you type, using TypeSafe's Jev, and puts you in charge of what happens next. A band above the prompt shows the gate score, the top skills with their probabilities, and the pick.

Using it

There are three modes:

• observe: Claude sees its normal skill listing, minus any skills you turned off, and loads whatever it picks.
• block: Claude sees no skill listing, and its Skill calls are refused.
• steer: Claude sees no skill listing, plus a hidden note naming Jev's pick. Only that skill can load.

Switch with /jev observe, /jev block or /jev steer, or the mode button in the band.

/jev opens a pane with a checkbox per skill. A skill you untick is left out of Jev's ranking, out of Claude's listing, and refused if Claude asks for it. /jev off turns the whole mod off: nothing is scored or sent, the band goes away and Claude picks skills as usual. /jev on brings it back. Your mode, on or off, and the unticked skills are kept across sessions.

Skills you type yourself as slash commands always run.

API key

You need a TypeSafe API key, from the TypeSafe console. Put it in the apiKey plugin option, which is kept in secure storage, or set the TYPESAFE_API_KEY environment variable. The option wins when both are set. Without a key, the band says so and nothing is scored.

Each prompt you type is sent to api.typesafe.ai along with your skills' names, descriptions, and the first 700 characters of the top three SKILL.md files.

How it works

It uses the two-request method from TypeSafe's skill suggestion cookbook:

1. You type. 600 ms after you stop, the draft goes to Jev. Nothing is sent to Claude. Drafts under 8 characters and slash commands aren't scored.
2. Jev ranks every skill. One Choice question over every skill's description, plus three Noul questions on whether the prompt needs a skill at all. If that gate is under 0.30, no skill.
3. Jev re-reads the top 3. This time with the opening 700 characters of each SKILL.md, and a fits question for each. If none fits above 0.30, no skill.
4. Your mode decides.

It finds skills in ~/.claude/skills and the project's .claude/skills, and learns plugin and built-in ones from Claude's own skill listing.

Limits

• Every prompt you type is sent to an outside service while you type it.
• In steer mode, the note tells Claude to ignore the pick if it doesn't fit, but only Jev's pick can load. If Jev picks nothing, no skill loads.
• The thresholds are fixed in the code, not settings.