The GRACE methodology, used for managing AI development projects, incorporates requirements, module graphs, and verification protocols. Despite clear semantic instructions, smaller AI models often struggle with implementation details, leading to inefficient token usage and repeated errors.
A controlled experiment tested the impact of adding few-shot examples to the GRACE framework. Models were tasked with generating React components based on identical specifications, with one group receiving additional verified implementation examples to follow as a structural guide.
The results indicated that providing one or two verified examples significantly improved the output of both small and large models. This approach helps agents adopt required patterns without needing to re-invent implementation methods, while maintaining strict adherence to project constraints.
This methodology is now implemented in GRACE via PR #63, which links examples to specific tasks, tests, and verification evidence. Further testing with models like GLM-5.3 and Claude Haiku 4.5 showed improved performance in identifying errors that were previously overlooked.
