Microsoft researchers show that retraining a simple instruction document can boost AI performance by about 23 points on GPT-5.5 without changing the model itself.

What happened: Microsoft and three Chinese universities developed SkillOpt, a method that treats instruction documents (called 'skills') as trainable objects. A separate optimizer model proposes small edits to the skill document, accepting only changes that improve performance on a validation set. When tested across six benchmarks with seven different AI models—including GPT-5.5 and smaller systems—SkillOpt outperformed or tied every comparison method, including handwritten skills and specialized optimization techniques.

Why it matters: Most AI improvement efforts today change the model's internal weights, which is expensive and time-consuming. SkillOpt keeps the target model completely frozen and only refines a plain text file of 300 to 2,000 tokens. This means the optimization happens once during training, while deployment stays simple—the model just receives the finished skill as context. Smaller models benefit as much as large ones, suggesting that procedural instructions can substitute for knowledge the model lacks.

What to watch: The finished skills stay compact, with improvements typically coming from just one to four accepted edits across four training epochs. The method depends on reliable automatic scoring, so it works best on tasks with measurable success (like spreadsheets and math), while open-ended tasks where success is hard to measure would need human or model-based judgments. Skills also transfer across models and environments—a skill trained on a larger model improves smaller ones, and a skill trained in one agent environment works in another.