GLM-5.2's Code Reviews Are Only as Good as Your Prompt
Kilo AI explores the limits of Zhipu AI's GLM-5.2 model in code review tasks, finding that its effectiveness depends heavily on prompt quality. While the model can catch some bugs and style issues, poorly crafted prompts lead to shallow or inaccurate feedback, making it a useful but constrained tool for developers.
Background
- Zhipu AI (智谱AI) is a major Chinese AI company, best known as the developer of the GLM (General Language Model) series. GLM-5.2 is its latest large language model iteration, competing with GPT-4, Claude, and similar frontier models.
- The post explores a key practical challenge with LLM-based code review: the model's output quality depends heavily on prompt engineering — how precisely you describe your project's architecture, standards, and review criteria. Poor prompts produce generic, shallow feedback.
- This reflects a broader industry debate: are LLMs useful for code review? The skeptical view is that they miss context-specific bugs and encourage rubber-stamping; the optimistic view is that they catch basic issues humans overlook. Zhipu's post lands between them — useful, but brittle.
- The piece also touches on a "positional bias" in LLMs: models tend to favor code placed earlier in a long prompt, which can skew review results if not accounted for.