Evaluator-Optimizer Agent Pattern
Independent PiSkill directory guide. The original prompt remains hosted by Anthropic Claude Cookbooks.
What does this prompt do?
Implements a generate-review-revise loop in which an evaluator applies explicit criteria and the worker improves output until it passes or reaches a safe limit.
Primary use case
Build an agent that iteratively evaluates and improves work.
Expected output
A bounded evaluator-optimizer workflow.
Inputs or variables
- task
- evaluation rubric
- iteration limit
Related prompts
Claude Frontend Aesthetics
Shows how concrete visual direction, layout goals, reference language, interaction details, and iteration constraints improve frontend generation beyond vague requests for polish.
Claude Text Classification
Builds a Claude classification workflow with clear labels, representative examples, structured outputs, ambiguity handling, and evaluation over realistic cases.
Claude Citation Workflow
Shows how to request, preserve, and validate citations so claims remain traceable to supplied source material and unsupported statements can be identified.
Claude Content Moderation
Creates a Claude moderation workflow from explicit policy categories, severity rules, examples, structured decisions, uncertainty handling, and evaluation data.
Claude Evaluation Builder
Builds an evaluation set and scoring workflow for a model task using representative inputs, clear criteria, repeatable execution, and analysis of failure patterns.
Claude SQL Query Generation
Demonstrates a practical SQL-generation workflow that supplies schema context, asks for the correct dialect, constrains changes, and checks generated queries before execution.