SCOREDevtools & code
Jev Skill Review
Kunal Jain@Kunal_Jain9
Jev is a decision model used to review a corpus of skills, providing ratings across rubric dimensions and identifying skills that may need distillation.
SignalReport: bottom-10 by mean score, top-5, needs_distill > 0.7, total input tokens, cost (tokens x $0.042/M), corpus mean.
I got an ivite for Jev from @typesafeai.
Its a decision model that costs $0.042/M input tokens. Free output. No prose.
The good part is you get $5 usage via the invite else you can buy usage on @OpenRouter.
I have a 216 skills corpus for zyta.dev and pointed Jev to review them.
Results:
216 skills.
6 rubric dimensions each.
~90 seconds.
$0.035 total.
Corpus mean: 3.46/5.
And the funniest result?
Our own skill-reviewer flunked itself: 2.56/5.
Jev rates. Code decides.
Copy this prompt to review your own skills.
You have a TypeSafe Jev API key (TYPESAFE_API_KEY). Rate every skill in my
agent-skill directory the way a strict reviewer would:
1. List all SKILL.md files under .agents/skills/ (or your skills root).
2. For each skill, POST api.typesafe.ai/v1/systemone with:
- state: {skill_name, dir, line_count, content}
- questions: one Score question per rubric dimension
(criteria: ["1: missing or broken","2: major gaps","3: adequate with
gaps","4: strong, minor gaps","5: exemplary"]) + one noul:
"needs_distill — too long, duplicated, or oversized for its job?"
- model: "jev-latest"
3. Run up to 5 in parallel, honor retry-after on 429, append each result
to .logs/jev-skill-ratings.jsonl (resumable: skip names already logged).
4. Report: bottom-10 by mean score, top-5, needs_distill > 0.7,
total input tokens, cost (tokens x $0.042/M), corpus mean.
Rules: scores are triage signals, not verdicts. Cross-check the bottom of
the pile with your deterministic skill audit before rewriting anything.
Jev rates; humans decide.