Jev engineering

Prompt engineering is about coaxing. Jev engineering is about specification. The builds that report sub-second latency and fractions of a cent share a handful of habits.

1. Keep the state small

Jev is priced per input token and its speed comes from doing one thing. The flight-search build that reported 7 seconds and $0.0039 for a full booking flow sends a compressed DOM, not the page. The email triage builds send the subject and first paragraph, not the thread.

2. Write criteria like a rubric, not a prompt

A Choice takes a map of option to description. The description is not decoration; it is the decision boundary. Compare:

"technical": "Technical"

with

"technical": "Bugs, errors, or an integration that used to work and stopped"

The second one is what the well-calibrated builds ship. Each option should be distinguishable from its neighbours by the description alone.

3. Use the probabilities

Every answer comes with probabilities and confidence. The fraud-detection and job-matching builds in the catalog act automatically above a threshold and route the rest to a human. That is the whole product: a fast, cheap first pass that knows when it is unsure.

4. Ask several questions per call

One request, many questions, one round trip. The builds that report the lowest cost per item batch their questions: route, score urgency, flag spam, all against the same state. Splitting them into separate calls multiplies latency for nothing.

5. Put it in the loop

Because a call is around a hundred milliseconds, Jev can sit inside loops that a chat model never could: every frame of a game, every step of a browser agent, every tick of a market feed. Several entries in Games and real time and Agents and browsers do exactly this. Design the loop so a single wrong decision is cheap to recover from, then let the loop run.

6. Report your numbers

The builds that get shared are the ones with a cost and a latency in the post. Measure end to end, say what you measured, and file it.