Skeleton Coding
An AI-assisted programming method in which developers supply a behavioral specification together with a high-level code format specifying program structure (the skeleton), thereby constraining the output space of AI coding agents and improving the accuracy and maintainability of generated code.
Timeline (last 90 days)
Skeleton Coding is a study whose core idea is that developers, while providing behavioral specifications, also specify the program structure (skeleton) in a high-level, code-like format to narrow the space of candidate programs.
The paper models AI-assisted programming as a control problem and proves that skeletons can exponentially reduce the number of tests developers need to provide.
On four models and three benchmarks, the best-performing skeleton specification reduced average test error by 48.4% compared with the best end-to-end behavioral specification.
Even when the agent can query the complete reference implementation, adding the skeleton still reduced average test error by up to 27.0% relative.
Skeletons significantly narrowed the average gap between small and large models, by about 52.4%.
Using skeletons reduced implementation cost by 15–26 times.
Skeletons reduced code erosion by an average of 24.7%.
The skeleton is not meant to replace tests but to complement them—the structure determines the candidate space, and the tests make the selection within it.
The paper's empirical evaluation covers four models and three benchmarks, but writing the skeleton itself requires cognitive effort from the developer.
All Facts (9)
Skeleton Coding is a study whose core idea is that developers, while providing behavioral specifications, also specify the program structure (skeleton) in a high-level, code-like format to narrow the space of candidate programs.
50%UnverifiedThe paper models AI-assisted programming as a control problem and proves that skeletons can exponentially reduce the number of tests developers need to provide.
50%UnverifiedOn four models and three benchmarks, the best-performing skeleton specification reduced average test error by 48.4% compared with the best end-to-end behavioral specification.
50%UnverifiedEven when the agent can query the complete reference implementation, adding the skeleton still reduced average test error by up to 27.0% relative.
50%UnverifiedSkeletons significantly narrowed the average gap between small and large models, by about 52.4%.
50%UnverifiedUsing skeletons reduced implementation cost by 15–26 times.
50%UnverifiedSkeletons reduced code erosion by an average of 24.7%.
50%UnverifiedThe skeleton is not meant to replace tests but to complement them—the structure determines the candidate space, and the tests make the selection within it.
50%UnverifiedThe paper's empirical evaluation covers four models and three benchmarks, but writing the skeleton itself requires cognitive effort from the developer.
50%