eval-build
Build evals you can trust
Own the workflow through a reproducible baseline, reusing trustworthy work already completed. Read the operating agreement. Discover first; ask only what inspection cannot resolve.
For bundled local tools and exact commands, read toolkit setup. Run the workspace installer for this skill; it preserves existing customized files.
Establish the current state
Inspect the app entrypoint, supported input shape, model/provider, tool wiring, and existing eval commands. Read any saved workflow state. Select one feature and a real product outcome. Determine whether it is in production, already traces requests, has a dataset, has human-reviewed failures, and has validated graders. Production status cannot always be inferred from deployment files; ask when uncertain.
Inventory actual evidence locations and access: logs, trace exports, support examples, datasets, annotations, test scripts, and existing dashboards. Record usable versus merely mentioned resources. Do not gather a generic metric wishlist.