| name | eval-designer |
| description | Design lightweight evaluations for prompts, skills, agents, and workflow changes: before/after test prompts, success criteria, negative cases, edge cases, regression risks, and manual or scripted eval plans. |
Eval Designer
Create the smallest evaluation that can detect whether a prompt, skill, or workflow change helped.
Workflow
- Define the target behavior, scope, and users.
- List likely failure modes and regressions.
- Build a compact eval set, usually 6–10 cases:
- positive cases where the behavior should happen
- negative cases where it should not trigger
- edge cases with ambiguity or conflicting goals
- regression cases for behavior that must stay unchanged
- Write expected behavior for each case before judging outputs.
- Choose a simple scoring rubric: pass/fail or 0–2 is usually enough.
- Propose a run plan: manual comparison, scripted harness, or session-log sampling.
- Explain how to interpret results and what would justify reverting.
Output
- Target behavior
- Failure modes
- Eval cases table: input/scenario, expected behavior, pass criteria
- Scoring rubric
- Run plan
- Decision rule: ship, iterate, or revert
Rules
- Use the output structure when helpful; otherwise adapt to the user’s requested format and keep the response proportional to the task.
- Default to 6–10 eval cases. Add more only when the user asks for broad coverage or the change is high-risk.
- Keep evals small enough to run before the change gets forgotten.
- Include non-trigger cases for skill/prompt activation changes.
- Prefer manual or semi-manual evals unless automation is cheap, repeatable, and likely to be reused.
- Avoid overfitting to one transcript; prefer representative prompts.
- Do not build a large benchmark unless explicitly requested.