handbooksWriting state and questions
Most bad Jev answers come from vague questions or noisy state. How to structure the evidence, phrase the question, and split big judgments into small ones.
- State is the evidence. Questions are what you want decided. Keep them separate.
- Use a JSON object for state when there is more than one piece of context. Named fields help.
- Ask small, snap-judgment questions. One decision per question.
- Jev reads questions literally. Write the exact condition you care about.
- Only include what the questions need. Extra text distracts it.
State: the evidence
State can be a string, a JSON object, or an array. For anything beyond a single message, use an object.
{
"subject": "Stripe sync broken",
"message": "My Stripe connection has failed for three days.",
"plan": "pro",
"previousTickets": 2
}Put everything a decision needs to compare into the same state. If the question is "does this order match the refund policy?", both the order and the policy go in.
Text only. Images, audio, and video are not supported, so describe them as text or structured fields first. English gives the best results. Other languages work, with lower accuracy.
Questions: small and literal
Think of each question as something a person could answer in one second after reading the state.
| Instead of | Ask |
|---|---|
| "Handle this incident" | "Which team should investigate?" |
| "Is this a good candidate?" | Four Scores: Python depth, leadership, system design, range |
| "Is this not a non-urgent message?" | "Does this need attention right now?" |
| "Is the invoice overdue?" | Compute the date difference in code, then ask about the rest |
Jev answers the question as written. It does not guess what you meant. If "urgent" means "customer is losing money right now" in your business, say exactly that in the instructions.
Split big judgments
A single question that tries to weigh five things at once gives a blurry answer. Five small questions give five sharp ones, and they run in parallel, so it costs you almost no extra time.
Then combine them in code, where you can see and change the logic:
# every score normalized to 0..1 by dividing by the top level
fit = (
0.4 * scores["python_depth"].score / 4
+ 0.4 * scores["system_design"].score / 4
+ 0.2 * scores["leadership"].score / 4
)If the ranking feels off, change a weight. You do not have to rewrite a prompt and hope.
Things Jev is bad at
These are the known weak spots for jev-1.13. Plan around them.
- Math and counting. Do arithmetic in code.
- Dates. Jev sees dates as text, not as ordered values. Compare them in code.
- Double negatives and long reasoning chains. Keep the question direct.
- Big, noisy state. Irrelevant text pulls answers off course. Trim it.
- Hostile input. Jev does not treat the state as possibly adversarial. Test with tricky inputs if users control the text.
- Opposite questions do not add to 1. "Is it spam?" and "Is it not spam?" are two separate answers. Ask one.
Keep math, filtering, and control flow in code. Give Jev only the judgment calls that code cannot make.
Where this page's facts come from▾
TypeSafe's pages on state, how to build with System One, composite scoring, and the jev-1.13 known issues. The incident example is from Vercel's What is Jev?.