evaluator-optimizer

Evaluator-Optimizer Loop

Generate, then have an independent critic score against an explicit rubric, then refine, until it passes or the budget is spent. One of Anthropic's five effective-agent patterns, and the one most starter kits omit. It is how you get measurably better output instead of one-shot output you hope is good.

When to use this

Invoke /evaluator-optimizer when both are true:

1. You can state acceptance criteria. There is a rubric, a spec, a set of must-haves, or a clear definition of done.
2. Iteration helps. A first pass is rarely the best pass for this kind of artifact: a specification, an API design, a security-sensitive change, an important document, a prompt you will reuse.

Do not use it when there is no clear evaluation signal, or when a single pass is obviously good enough. Adding a critic loop to a trivial task is the over-engineering this kit warns against (see .claude/rules/ai-orchestration-decision-gate.md).

The loop

1. RUBRIC      Define explicit, checkable acceptance criteria up front.
2. GENERATE    A generator produces the artifact against the rubric.
3. EVALUATE    The `evaluator` agent (frontier, read-only) scores each criterion: PASS / FAIL + concrete gap.
4. DECIDE      All blocking criteria pass        -> accept.
               Gaps remain and rounds remain     -> feed the evaluator's gaps back to the generator, go to 2.
               Budget spent and gaps remain       -> stop, hand back with the open gaps named (do not pretend done).

Keep it to 1 to 2 refinement rounds by default. More than that usually means the rubric is wrong or the task is mis-scoped, not that another round will help.

Roles

• Generator: the implementation or authoring agent (or the main thread). Produces and revises.
• Critic: the evaluator agent (.claude/agents/evaluator.md). Read-only, frontier model, scores against the rubric, never edits. Keep these two separate; a generator grading its own work is not an evaluation.