Playground
Ask Jev something narrow.
State on the left, questions under it, the official JSON on the right. Run sends that JSON to api.typesafe.ai/v1/systemone with the key in this browser.
1. State
JSON object or plain text. Every question sees this same state and is scored independently.
2. Questions
How to read a run
The bars are the product
After a live call, each question comes back as its own block. A choice shows the winning label and a bar per option. A noul shows one probability: that is the whole answer. A score shows the expected level and the weight on each rung you defined.
Presets
- Support triage — team, urgency, frustration. Speculative on purpose: you can ignore urgency if the team confidence is already too low to route.
- Tool guard — allow, confirm, review, or deny, plus whether the call is hard to undo. The tool does not run.
- Task router — proceed fast, review deeply, split, or block.
- Citation check — the evidence is only the text in the state. Nothing is fetched.
- Invoice coding — one account, a split, or a human. The amount stays a number for your ledger.
- Context keep — three snippets, three independent choices. Kept text stays verbatim because the model never rewrites it.
Editing questions
Ids are for your code. Instructions are what the model reads. For a choice, an empty description sends null and the option name stands alone — fine when the name is already plain, worse when two options overlap. For a noul, fill both “yes means” and “no means” when the boundary is subtle. Leave both empty when the question is already sharp.
The JSON on the right is the body of POST /v1/systemone. Copy it into the Python SDK or the preset API if you would rather not build the questions by hand.
The key
It lives in this tab’s session storage. Closing the tab drops it. The server attaches it as a Bearer token, calls TypeSafe, and does not write it to disk. A wrong key comes back as the upstream error, usually 401. There is no shared demo key: a run spends your quota.