Microsoft’s SkillOpt approach reportedly boosts GPT-5.5 by optimizing a Markdown-based skill file instead of changing the model weights. That makes the result notable because it treats structured instructions as a trainable asset, not just static documentation.
The work fits a broader trend in AI engineering: performance gains increasingly come from scaffolding, tools, memory, and task-specific procedures around foundation models. If reliable, these methods could offer teams a cheaper and more controllable alternative to fine-tuning for many workflows.