Question 1

Is a fraction of a cent per turn actually cheaper than listing skills?

Accepted Answer

It depends entirely on your library size, and the README gives both sides. A thirteen-skill listing costs about 1,350 tokens of frontier input per turn, which the author puts at $0.007 per turn at $5 per million and notes does not scale. Against that, a 23-skill library costs $0.0002 per turn and a 624-skill library $0.004. The crossover is where BM25 prefiltering starts, at 1,200 skills by default.

Question 2

What does BM25 prefiltering cost me in accuracy?

Accepted Answer

The table reports the prefiltered row, 3,024 skills narrowed to 400 sent, at 1.4-1.7 s and $0.0026 per turn, with precision exact in every row by the author's reckoning. The mechanism is a plain lexical ranking over names and descriptions in process, in milliseconds. The README argues for it over a model-driven category stage, which it tried and rejected because it dropped a skill whose folder was named `canary`.

Question 3

Do my skills disappear, or just move out of the prompt?

Accepted Answer

They move. `migrate` shows what would move and `--apply` moves ~/.claude/skills into the plugin's library, and you can also just drop SKILL.md folders there yourself, nested folders included. The observer scans those folders and injects a body only when it scores at or above the threshold, so a skill is off the prompt until something makes it relevant.

Question 4

A skill of mine never loads. How do I find out why?

Accepted Answer

Use `node dist/cli.js ask <prompt>`, which runs the observer on a prompt and prints the decisions, and `status` for the log lines. The README's advice is direct: if Jev returns 0.40 when you expected 0.95, the description is the thing to fix. The tunable sits in ~/.jev-skills/config.json, where threshold, maxSkillsPerTurn, injectBudgetChars and questionStyle all live.

Question 5

How do I install it for Claude Code, and for Codex?

Accepted Answer

Claude Code: claude plugin marketplace add eran-broder/jev-skills, then claude plugin install jev-skills@jev-skills --config typesafe_api_key=. The key is declared sensitive plugin config, stored in Claude's credential store and handed to the hook as an environment variable, so it never touches a settings file. Codex: clone, then node dist/cli.js setup and node dist/cli.js install-codex, which writes four hooks into ~/.codex/hooks.json with absolute paths and asks for trust once.

Question 6

What are the hard limits, and what happens at scale?

Accepted Answer

Jev's limits bind, not the model's: 64k tokens per request, which is about 400 skills at 150 tokens each, plus 250k tokens per second and 1,200 requests per minute per account. Direct fan-out over 3,000 skills spends 461k tokens per prompt, so a whole account gets about 30 prompts a minute. The hook chunks requests automatically and sends chunks in parallel, and latency follows tokens per request more than request count.

Question 7

What is tunable, and what is the license?

Accepted Answer

Everything in ~/.jev-skills/config.json is optional: threshold 0.6, maxSkillsPerTurn 3, injectMode full or pointer, injectBudgetChars 24000, observeToolResults, toolResultMinChars 200, questionStyle full or slim where slim moves the shared rubric into the state for 36 percent fewer Jev tokens, prefilterAbove 1200, prefilterKeep 400, timeoutMs 6000. Injection happens once per session and compaction resets it. MIT, and dist/ is committed so hooks run with no install step.

Jev Skills

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community