Building with Jev plugin
A skill that teaches an agent to write programs for TypeSafe Jev: Choice, Score, Noul and Box questions, state, thresholds
- 167 ★ on the repository
- Last commit
Minimal access
Why this level
- The skill only explains Jev techniques to the agent, while paid model calls run under your own TypeSafe key
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/building-with-jevDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The skill only explains Jev techniques to the agent, while paid model calls run under your own TypeSafe key
This is third-party code. Review the repository files before installing.
What it does
The `jev` skill is written for Claude Code, Codex, Cursor and other agents and helps write and improve programs that call Jev, the System One model from TypeSafe. Jev does not reason in steps and does not generate text: it reads one `state`, answers every question in the request independently and in parallel and returns a probability distribution over the answers you defined. Code makes the decisions. The skill covers five topics: designing questions of four types (Choice for a pick from a set, Score for a position on a scale, Noul and Box), state structure including images, composing answers in code, confidence thresholds and diagnosing questions that answer wrong or with low confidence. The working order has seven steps: list the decisions, one question per judgment, pick the primitive, build the smallest state, send one shared request, combine in code and test on labeled examples. The guide targets the `jev-1.14` and `jev-1.13` models. Image and Box questions are experimental in the TypeSafe API and the skill marks them that way. Details live in `skills/jev/references/images-and-box.md`, which the agent reads only when needed.
Who it is for. Developers who build sorting, scoring or next-step selection on the TypeSafe API and ask an agent to write the questions and thresholds.
Good fit when
- The task "analyze this message and decide what to do" is cut into small questions that Jev rates in a second
- A question answers wrong or with low confidence and you study `probabilities` on the misses
- The decision is composed in code from answers: branches, weights and confidence thresholds sit in one file
- Tuned thresholds are pinned with a versioned model name instead of the `jev-latest` alias
Not a fit when
- You need text generation or step-by-step reasoning: Jev gives probabilities over defined answers and writes nothing
- The main input language is not English: Jev accuracy is lower in other languages and thresholds need more tuning
- You need a reference for the TypeSafe API itself: the company own skill in typesafe-ai/skills covers it
Example request
Design Jev questions to sort incoming emails by urgency and topic, move thresholds into a separate file and prepare labeled examples for testingLimitations
The skill helps design questions and does not replace model access: calling Jev needs a TypeSafe key and paid requests. Recommendations apply to `jev-1.14` and `jev-1.13`, and the jaggedness page should be reread for later versions. Image and Box questions are experimental in the API. English gives the best accuracy and other languages, including CJK, work worse. The author notes that agents write weak questions, so wording and threshold constants should live in one file and be edited by hand.
How to disable. In Claude Code run `/plugin uninstall jev@building-with-jev`. For a skill installed through the skills CLI, delete the `jev` folder from the agent skills directory.
FAQ
How do I install Building with Jev in Claude Code?
Run `/plugin marketplace add dbreunig/building-with-jev-skill`, then `/plugin install jev@building-with-jev`. The skills CLI installs it with `npx skills add dbreunig/building-with-jev-skill`, which works with Claude Code, Codex and Cursor.
How do Choice and Score differ?
Choice picks one option from an unordered set and returns `choice`, `probabilities` and `confidence`. Score gives a position on a scale described in steps and returns `score` and `legend`. Code branches on Choice and applies a threshold or rank on Score.
What makes a good Jev question?
One a knowledgeable person answers in a second given the right context, such as "does this message convey urgency?". A question that weighs two properties should be split into two.
Which model version should I specify?
For stable answers pin a versioned name such as `jev-1.14.0`: the `jev-latest` and `jev-preview` aliases move with releases. Log `response.model`.
How is this skill different from the official TypeSafe skill?
The official TypeSafe skill covers the API and helps find places to use it. This one focuses on writing questions, composing answers and diagnosing bad questions.
Discussion
No questions or reviews yet.
Related
PluginsAround 200 skills for Claude Code on bioinformatics, chemistry and biostatistics, from RNA-seq analysis to molecular docking
Probabl skills for machine learning experiments: a scikit-learn and skrub pipeline, evaluation, audit and a link to Skore Hub
Live data from 78 platforms as JSON through one key: skills teach an agent to call the Scavio API for Amazon, Airbnb and more
43 skills for Claude Code along the research cycle: surveys, experiments, topic models, reporting and manuscript review
A skill that picks a scientific chart from nearly 100 templates for CSV or Excel data and outputs PNG, SVG, PDF and Python code.
StatsClaw assembles nine Claude Code agents to build statistical packages from a paper or formulas with isolated pipelines.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Building with Jev (this card) | no | 167 | |
| TypeSafe Agent Skills | no | 2,895 | a skill for connecting TypeSafe and System One questions, while this focuses on question wording, thresholds and diagnosis |
| Jev Skills | no | 668 | five skills for sorting, source checks, rubric scoring and step selection, while this is one skill for writing the Jev programs themselves |
README in short
The README lists the five skill topics, names the target models jev-1.14 and jev-1.13 and notes that image and Box questions are experimental. Installation as a Claude Code plugin, through the skills CLI or manually follows.