skill-construire-evaluations

To create an evals.json file for prompt testing, generate nominal, edge, incomplete, and neighbour prompts that separate trigger and non-trigger cases, then express expected behavior with precise assertions for regression coverage.

Building regression coverage for weak prompt outputs involves documenting triggering, non-triggering, and edge case prompts in an evals.json file with explicit assertions to guard against future changes and false positives.

Writing evaluation cases that separate trigger and non-trigger prompts requires generating nominal and neighbour inputs, then recording expected behavior assertions for both triggering and non-triggering scenarios in an evals.json file.

You need to generate edge case evaluations when a new capability requires regression coverage or when you are refining a target skill after encountering false positives or weak outputs.

No, this evaluation approach cannot rewrite a skill with weak outputs; it is designed strictly to generate evaluation cases and an evals.json file to test triggering behavior and document edge cases.

Yes, generating an evals.json regression suite requires referencing motifs-evaluations to construct comprehensive trigger, non-trigger, and edge case assertions for the target skill.