How to Get Started With Jev (Without Cargo-Culting the Hype)

Here's the thing about getting started with Jev: the syntax is the easy part, and it's documented by the people who actually control it at docs.typesafe.ai. We're not going to reproduce API details on a third-party site where they'd rot within a week of any update; launch-week APIs move fast, and a tutorial with stale endpoints is worse than no tutorial. For orientation while you read those docs: the ecosystem's clients and directories consistently describe the model as typesafe-ai/jev, called through the Vercel AI Gateway, with an evaluation surface named experimental_evaluate and official SDKs (Python first) plus community clients in TypeScript, Swift, and Elixir. Treat every one of those details as the ecosystem's report, and the official docs as the ruling.

What we can give you is the part the official docs assume you already have: the mental model. Because six days into cataloging 600+ builds, the pattern is unmistakable: the builders shipping impressive things aren't better programmers, they're better question designers. That's the skill. Here's how to get it.

Step 0: Unlearn the chat reflex

If you've spent two years prompting chat models, your instinct is to hand the model context and ask it to figure things out. Jev punishes that instinct. It doesn't converse, doesn't explain, doesn't handle "analyze this and tell me what you think." It answers constrained questions: a label from your list, a score on your scale, a yes/no on your criterion.

The reframe that unlocks everything: stop asking the model to think; ask it to rule. A judge, not a consultant.

Step 1: Decompose your task into verdicts

Take the job you want automated and keep splitting it until every piece is a question with a small, closed answer set. In pseudocode (illustrative pattern, not Jev syntax; real request format is in the official docs):

# PSEUDOCODE — the shape, not the API
for email in inbox:
    category = decide(email, "Which: support | sales | billing | spam?")
    urgent   = decide(email, "Action needed within 4 hours to prevent customer harm? yes/no")
    route(email, category, urgent)

One real constraint worth knowing before you design: choice sets cap out around 255 options per question (community libraries like jev-tree exist to recurse past it), so taxonomies bigger than that get split into staged questions.

Notice what makes the second question good: it's operational. "Is this urgent?" invites vibes; "action within 4 hours to prevent customer harm" invites a ruling. Every great build in the triage category is really a stack of questions written this carefully. The Jev Engineering guide is our deep dive on the craft; treat it as required reading after this page.

Step 2: Pick a first project with receipts to copy

Don't invent; imitate, then mutate. Three starter lanes, each with working precedents in the directory:

Inbox or ticket triage (weekend-sized, immediately useful): classify a few hundred of your own emails, compare against your judgment, tune the questions. Precedents: 500 emails for 3.5 cents, the full email-triage pattern.

A judge for something you already eval by hand (highest career ROI): PR descriptions, AI outputs, content drafts. One builder asks 61 questions per draft for $0.0004 (build); the LLM-as-a-judge guide covers the pattern and its failure modes.

An agent loop (most fun): browser automation where Jev picks each next action; the browser-agent builds report ~$0.001 to $0.004 per completed task.

Step 3: Measure before you trust

The directory's whole ethos is receipts, so generate yours. Before wiring any verdict to a real action: run 50 to 100 cases, score them against your own answers, find where the model disagrees with you, then decide whether the fix is a better question (usually) or an escalation tier (for the genuinely hard slice). The fraud build's cascade design (Jev rules on everything, uncertain cases go to a bigger model) is the pattern to copy the moment accuracy matters. And log costs from call one; the numbers are small enough that people forget to check, which is how "cheap" quietly becomes "unmetered."

Step 4: Ship it and claim your receipt

Publish the numbers: what it did, how fast, what it cost. Then submit it. Half the entries in this directory started as "small weekend thing, probably nothing," and the ones with honest receipts are the ones that travel.

Frequently asked questions

Where are the official Jev docs?

docs.typesafe.ai, run by TypeSafe AI. API syntax, access, auth, and rate limits are theirs; this independent site covers patterns, builds, and economics.

Do I need ML experience to use Jev?

No. If you can write an if-statement and a clear sentence, you can use a decision model. The learnable skill is question design, which is closer to writing good acceptance criteria than to machine learning.

What should my first Jev project be?

Something with volume you personally understand: your inbox, your team's tickets, drafts you already review. You need your own ground truth to check verdicts against; novelty can come second.

What are Jev's rate limits and pricing tiers?

TypeSafe's to state, and launch-period details change fast; check the official docs rather than any third-party snapshot, including ours. For what real workloads have cost builders in practice, see what builds actually cost.

What's the most common beginner mistake?

Vague questions. "Is this good?" produces coin flips; "does this post make a specific claim a reader could verify?" produces signal. Second most common: trusting verdicts before scoring them against 100 known cases. Do the boring calibration.