Meta's Structured Prompting Lifts Code Review Accuracy to 93% in Repository Tasks

Meta researchers have developed a structured prompting method that significantly improves large language model performance on code review and repository-scale tasks, achieving accuracy as high as 93% in certain cases by replacing expensive execution sandboxes with guided reasoning frameworks.

The technique addresses a persistent bottleneck in deploying AI agents for bug detection, patch verification, and code review: the computational cost of setting up dynamic execution environments for every repository. By using semi-formal prompt templates to structure LLM reasoning, the approach sidesteps execution overhead while reducing hallucinations that plague unstructured inference.

Yet the method shows clear limits. When researchers tested it on Sonnet-4.5 for code question-answering tasks where the model already achieved 85% accuracy, the structured templates yielded no additional gains. The findings suggest that prompt engineering retains value in specific contexts even as some industry voices declare the discipline obsolete.

Meta has released the prompt templates publicly, enabling immediate integration into developer workflows. The disclosure comes as OpenAI reported crossing $2 billion in monthly revenue—quadruple the growth rate of Alphabet and Meta during comparable stages—while expanding its Codex platform into a flagship coding agent and launching GPT-5.4 with gains in workflow performance.

(The legal industry is simultaneously racing to acquire AI expertise, with SurePoint Technologies' 2025 report finding firms prioritizing lateral hires specializing in artificial intelligence and seeking "outward marketability" through technology competence, particularly among midsize practices hoping to compete with larger rivals.)

The code review breakthrough arrives amid heightened scrutiny of AI development practices. Anthropic is managing fallout from a leak of code underlying its Claude AI agent, while industry observers debate whether tools like OpenClaw—now being combined with AI coding platforms by some startups to "fully automate developers' jobs"—signal commoditization or a new phase of capability concentration among leading labs.