code-review

name	code-review
description	Use when reviewing AI-generated code for architectural quality, design patterns, and engineering practices

Code Review for Assessment Evaluation

You are reviewing code produced by an AI coding agent. Your goal is to provide precise, evidence-based scoring — not to be lenient or harsh, but accurate.

Review methodology

1. Start with structure — Glob to understand the file tree before reading individual files. The shape of the codebase tells you about architectural decisions.

2. Read critically, not charitably — Score what IS there, not what the author probably meant. If a pattern is half-implemented, score it as half-implemented.

3. Trace the domain model — Follow the flow from entry point to persistence. Look for:

   • Are domain concepts explicit types or buried in primitives?
   • Do boundaries between modules/layers exist and hold?
   • Is business logic in the domain or scattered across infrastructure?

4. Check encapsulation — Look for:

   • Public fields that should be private
   • Getter/setter pairs that expose internals
   • Domain objects that are just data bags with no behavior
   • Invariants that are enforced externally rather than internally

5. Evaluate test quality — Tests that merely exist are not enough. Check:

   • Do tests verify behavior or just call methods?
   • Are edge cases and failure modes covered?
   • Do test names describe the scenario being tested?
   • Are tests testing the unit or the framework?

6. Look for anti-patterns — Common problems to flag:

   • Anemic domain models (logic in services, entities are just DTOs)
   • Leaking abstractions (domain depends on infrastructure types)
   • God classes or methods doing too many things
   • Copy-paste with minor variations instead of proper abstraction

Scoring principles

• Evidence required — Every score must cite specific files, classes, or code patterns. "The code generally looks good" is not evidence.
• Calibration — A max score means excellent, not merely acceptable. Reserve top scores for genuinely well-crafted code.
• Partial credit — If the agent solved the core problem but cut corners on secondary concerns, reflect both in the score and reasoning.
• Zero scores are valid — If a dimension was completely ignored by the agent, score 0 with explanation.