Building with Jev

What it does

Building with Jev is an agent skill by Drew Breunig: one SKILL.md file that teaches Claude Code, Codex, Cursor, and similar agents how to write and improve programs that call Jev. It covers:

• Question design. Choosing between Choice, Score, and Noul, writing instructions that ask one exact thing, and writing criteria that point the same way as the question.
• State. Sending only the fields the questions need, and doing arithmetic, dates, and counting in code.
• Composition. Speculative fan-out in one request, confidence-gated routing, composite scores, taxonomy walks, and extraction with candidate lists.
• Diagnosis. A table that maps symptoms, such as Nouls hovering near 0.5 or Scores clustering in the middle, to likely causes and fixes, plus rules for revising questions safely.

It ends with a twelve-point checklist and links to the TypeSafe docs pages it draws on.

Setup

In Claude Code:

/plugin marketplace add dbreunig/building-with-jev-skill
/plugin install jev@building-with-jev

It also installs with the skills CLI (npx skills add dbreunig/building-with-jev-skill) or by copying the folder into your skills directory. The agent loads it when a task involves Jev, or you can call it with /jev.

What we checked

A skill is only as good as its facts, so we checked the technical claims against TypeSafe’s documentation, which we keep a daily snapshot of.

• The limits are right. The 64k-token request limit, with 32k for the state plus the longest question, matches the models page. So do Score’s range of two to ten levels, and the fact that a Noul returns no confidence.
• The subtle points are right too. Question IDs aren’t sent to the model, and each Score level is judged on its own without its number or neighbors. Both statements match the docs. The instruction keys it recommends (question, focus, inspect, compare, and a structured field record) all appear in the docs’ own examples.
• The patterns follow the docs. The confidence floors (about 0.6, rising to 0.85–0.9 for risky actions) and the composite-score normalization by the number of levels minus one match TypeSafe’s pattern pages.
• The code example runs. We ran the skill’s Python routing example unchanged with TypeSafe’s official SDK (0.7.0) and a real key. It executed without errors, returned the three expected answer types, and called exactly one routing function.
• The plugin manifest validates. Claude Code’s plugin validator accepted its marketplace manifest.

We found nothing inaccurate. Where the skill adds its own advice, such as putting the neighboring case in the criteria of the side it belongs to, it’s clearly consistent with the docs rather than contradicting them.

Permissions and data

It’s plain text. The skill has no code, makes no network requests, and needs no keys. It only shapes what your agent writes. The programs your agent builds will call TypeSafe with your key, as usual.

Maintenance and alternatives

A single commit on September 17, two days after Jev launched, with no open issues. There’s no license file, so it’s unclear whether you may copy or adapt the text beyond installing it. The skill targets jev-1.13 and says to recheck TypeSafe’s jaggedness page when the model changes.

TypeSafe publishes an official agent skill too. It’s MIT licensed, installs the same way, and is shorter. It starts by telling the agent to read the live docs. Drew Breunig’s skill instead puts the key guidance, the diagnosis table, and the checklist directly in the file. We didn’t compare them line by line, but both draw on the same docs, and it’s reasonable to install both: the official one for currency, this one for its concrete fixes.

Verdict

Building with Jev is an accurate, tightly written distillation of how to design Jev questions, and its diagnosis table is a genuinely useful page to keep open when a question misbehaves. Install it if you build Jev programs with a coding agent, and read it even if you don’t. Just remember it describes jev-1.13, and ask the author about a license before reusing the text.

For turning its threshold advice into numbers from your own data, see jevcal and our confidence thresholds guide. For letting your agent call Jev directly, see System One Connector.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 19, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.