Choose a setup you can evaluate.
An unfiltered or uncensored label is not a measurement of coding quality. Choose a model and endpoint, then evaluate a small task whose answer you can verify. Behavior and restrictions depend on the selected provider, model, client, and tools.
- 01
Establish the setup
Confirm the model identifier, endpoint, credential requirements, and supported context in the selected provider documentation. Understand whether requests leave your machine and whether connected tools contact other services. Do not infer data handling from the appearance of a local desktop application.
- 02
Use a representative task
Ask the assistant to explain a small function, identify its callers, and suggest a relevant check. Compare cited files and reasoning with the repository. A model that answers a general coding puzzle well may still struggle with your project's conventions or incomplete context.
- 03
Evaluate the whole iteration
Look at the correctness of the result, unnecessary edits, response to feedback, and quality of verification. Use the same task and starting state when comparing setups. Avoid fixed speed, cost, or privacy assumptions; check the current terms and behavior of the route you actually select.
A useful finish
You understand the endpoint route and have inspected a representative response against real project evidence.
A small model evaluation
Read only. Explain [function] and locate its callers and closest test. Cite repository evidence and identify uncertainty. Suggest one meaningful verification. Do not edit files, install packages, or infer project behavior from naming alone.