Small corrections. Clearer thinking.

A little nudge.
Better judgement.

Jev-based concept linting to nudge your LLM to improve its judgement.

The code compiles. The tests pass. But does the commit say anything? Does that comment explain why? Sometimes the next improvement needs a different kind of check.

Open source MIT licensed Built on Jev

Fig. 01 — A small course correction
First draftA little clearer.a useful question
Same intent. A more thoughtful result.
A second look at

Commits / Code / Comments / Tests / Writing

01 / The nudgements

Looks fine.
Look again.

Small findings from real open-source code.
A before, a nudge, and a more considered after.

Name what actually changed.

Source: ansi-regex

Before

Improve performance

After

fix: stop unterminated OSC strings consuming later escapes

The nudgement

The subject is generic. Name what actually changed.
Commit evaluator

Recorded evaluator responses. Proposed edits, not upstream changes. Read the full examples · Source licenses.

02 / Under the hood

Lint the idea.
Not just the syntax.

A formatter knows where the semicolon goes. nudgement asks whether the thought holds up.

  1. 01

    Give it context.

    A file, a diff, a commit message. nudgement gathers the relevant context and runs exact checks.

    draft + context
  2. 02

    Ask a focused question.

    TypeSafe’s Jev returns structured judgements. Does this message match the diff? Would this test catch a broken implementation?

    questions → judgements
  3. 03

    Make a better call.

    nudgement turns those answers into findings. You or your coding agent decide what to change, then review the next draft.

    findings → revision

A nudge is a reason to take another look. It isn’t proof of correctness, and it doesn’t rewrite your code for you.

03 / Put it to work

Give your agent
a second thought.

Run nudgement before a commit, or give your coding agent the included skill. Keep the useful findings. Refine. Repeat.

One more look before you commit
# From your nudgement checkout
bun evaluate.ts /path/to/your-repo \
  --staged \
  -m "fix(cli): handle --help"
Needs Bun, a clone of the repo, and JEV_API_KEY.
Setup instructions

Reviews send selected text and context to Jev and may incur API charges. The examples on this page use saved responses.

Good judgement is a practice.

Keep nudging·

A small open-source experiment by Nicholas Westby.

Read the MIT license. Make it your own.