Building with Jev (Claude skill)
github.com/dbreunig/building-with-jev-skill- Category
- Developer Tools
- Rank
- No. 1216Tools index
Previous survey · No. 1234 ·
- Type
- TOOL
- Use case
- Coding
- Interfaces
- Agent Skill / Plugin
- Builder
- @dbreunig
- GitHub
- 164 stars
- Date
About
This repository packages a Claude Code agent skill for writing and improving programs that call Jev, TypeSafe's System One model. It covers question design (Choice, Score, Noul question types), state structure, composing answers in code, setting confidence thresholds, and diagnosing questions that return wrong or low-confidence answers, targeting model version jev-1.13. It installs through the Claude Code plugin marketplace or the skills CLI, and Claude loads it automatically when a task involves Jev.
What it does
A plain-text guide that teaches Claude Code, and other coding agents, how to write good programs against Jev, a judgment model that answers many small questions about one input and returns probabilities rather than text. The guide tells the agent how to break a fuzzy decision into one-property questions, pick the right answer type, keep the input small, ask everything in one request, and route on confidence in code. A separate page covers images and bounding-box questions, loaded only when a task uses images.
Why it's ranked here
It is a narrow tool done carefully. The guidance is concrete and checkable: Choice questions accept up to 255 options, Score levels run two to ten, the input and all questions share 64k tokens, and starting yes and no bands sit at 0.7 and 0.3. It separates itself from TypeSafe's own skill by focusing on writing and fixing questions rather than API coverage. The value depends entirely on whether you build with Jev; outside that it does nothing.
What's good
The advice targets the mistakes people actually make with classifier-style models: questions that weigh two properties at once, yes-or-no questions phrased so a high value means no, and a single 0.5 cutoff that flips between runs. It tells the agent to pin a versioned model once thresholds are tuned, since aliases move with new releases. It warns that text inside the input can steer answers and should be tested before deployment. Experimental features are labeled as such, and the image page lists exact size and format limits.
Tradeoffs
It is documentation for a single vendor's model, so it ages with that vendor's releases; it targets two specific Jev versions and says limits may change later. It contains no code, tests or examples you can run, only prose and short snippets. Image and box guidance covers features TypeSafe itself calls prototype-grade. The readable files do not state a license. Its value rests on the accuracy of claims drawn from TypeSafe's docs, which the guide cites but does not verify with tests.
How to use it well
Install it if you or your agent write code that calls Jev, through the Claude Code plugin marketplace, the skills command-line installer, or by copying the folder into your skills directory. Claude picks it up when a task mentions Jev, or you can call it by name. Use it alongside TypeSafe's own skill, which covers the API surface. Keep its suggested habit of storing questions and thresholds together, and test changes against labeled examples before shipping.
Technical notes+
skills/jev/SKILL.md carries frontmatter name jev and a trigger description covering Choice, Score, Noul and Box; it documents model aliases jev-latest and jev-preview pointing at jev-1.14.0, token limits of 64k shared and 32k per question, a TypeSafe-Danger-Zone-Long-Context header raising both to 262k, and a 32 MiB request body with 64 levels of nesting. skills/jev/references/images-and-box.md requires a TypeSafe-Danger-Zone-Images header (400 without it), PNG or JPEG at 280 to 1568 pixels per side, 16 MiB per image and 32 MiB total, and notes Box needs jev-1.14. .claude-plugin/plugin.json names the plugin jev by Drew Breunig; .claude-plugin/marketplace.json registers the marketplace building-with-jev with source ./. README.md documents three install routes.
Observed
- Format
- Agent skill: Markdown instructions with a reference page
- Install surface
- Claude Code plugin marketplace, skills CLI via npx, or manual copy
- Agents
- Claude Code; the skills CLI also targets Codex, Cursor and others
- Target models
- Jev 1.14 and 1.13 from TypeSafe
- Executable code
- None; prose guidance with illustrative snippets
- License
- Not stated in README.md or .claude-plugin/plugin.json
- Experimental scope
- Images and Box questions marked experimental (skills/jev/references/images-and-box.md)
Read from README.md, .claude-plugin/plugin.json, skills/jev/SKILL.md, skills/jev/references/images-and-box.md, .claude-plugin/marketplace.json.
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.