| name | agentic-eval |
| description | Evaluate and improve AI workflow outputs with small goldens, rubrics, and repeatable checks; use when tuning skills, prompts, instructions, or agent behavior. |
| category | research-iteration |
| kind | atomic |
| phase | analysis |
| requires_write | true |
| requires_slurm | false |
| requires_benchmark_artifacts | false |
| delegates_to | [] |
| output_schema | skill_run_summary.v1 |
Agentic Eval
When to use
Use this skill when the target is an AI workflow artifact (skills, prompts, rubrics, issue text,
review guidance, etc.) rather than core product code.
Read First
AGENTS.md
docs/dev_guide.md
docs/ai/awesome_copilot_adaptation.md
docs/code_review.md
Workflow
-
Define the artifact and objective
- State the artifact, expected behavior, and what a pass/fail result looks like.
- Set a small representative evaluation set.
-
Build a rubric
- Use 3–6 concrete criteria.
- Mark critical failures separately from quality preferences.
-
Create a baseline
- Score the current artifact against the rubric before editing.
-
Improve one dimension at a time
- Make a targeted change.
- Re-run the same evaluation.
- Keep the change only if improvement is clear and reproducible.
-
Capture proof
- Compare baseline vs post-change with the same rubric and commands.
- Keep the result compact: what improved, what regressed, what remains unknown.
Proof and Guardrails
- Always preserve baseline and post-change results.
- Reuse the same corpus for all iterations.
- If changes drift from the target behavior, revert and document the false regression.
- Prefer
autoresearch when the task has repeated measurable benchmarks.
- If benchmark evidence is required, treat failed/invalid runs as inconclusive, not proof.
Output
Report:
- artifact inspected,
- rubric and dataset,
- baseline and final score,
- keep/discard decision with rationale,
- open questions.
Guardrails
- Stay within the skill scope declared in
.agents/skills/skills.yaml.
- Prefer repository scripts and canonical docs before ad-hoc commands.
- Record blockers and validation gaps instead of overstating completion.