01
Load tool definitions
Anthropic tools, OpenAI tools, or a bare array of name + description + schema. Format is auto-detected.
02
Ergonomics lint
Static checks for the ambiguities that make models pick the wrong tool or invent arguments.
Nothing loaded yet. Parse definitions on the left and the lint runs automatically.
03
Play the model
Write yourself a task, then act as the model would: choose a tool and fill the arguments using only what the definitions say. If you need to peek at your backend code to fill a field, the model is worse off than you - it cannot peek.
This becomes the "task" field of the exported eval case.
Load definitions in step 01 first.
Select a tool to generate its argument form from the schema.
04
Transcript
Chronological record of every call with args, validation outcome, and your notes on what confused you.
No calls yet. Submit a call in step 03 and it lands here with its validation outcome.