Microsoft researchers partnered with three Chinese universities to develop SkillOpt, a novel optimization technique that improves AI agent performance through instruction refinement rather than model retraining. The method works by iteratively enhancing Markdown-formatted instruction documents using principles borrowed from traditional neural network training.

The approach delivers measurable gains without touching the underlying model weights. On procedural tasks, SkillOpt improved GPT-5.5 performance by approximately 23 points. The optimized instruction files proved portable across different models and agent environments, functioning effectively with both Codex and Claude Code.

This represents a shift in how organizations approach AI performance tuning. Instead of expensive full model retraining or fine-tuning, teams can now optimize the instruction layer. The Markdown format keeps the process transparent and human-readable, making it accessible to practitioners without deep machine learning expertise.

The portability across models has practical implications for enterprises. A single optimized instruction set created for one model can be reused when migrating to different systems, reducing redundant optimization work. This works because the improvements target instruction clarity and structure rather than model-specific quirks.

SkillOpt aligns with broader industry trends toward post-training optimization. Rather than constantly retraining larger models, researchers focus on extracting better performance from existing systems through smarter prompting, retrieval systems, and now instruction optimization. The method scales efficiently since Markdown files require minimal computational resources compared to full model training.

The collaboration between Microsoft and Chinese academic institutions signals continued cross-border AI research despite geopolitical tensions. The work suggests practical tools for enterprise deployment where instruction optimization can be continuously applied as teams learn what works best for specific tasks.

For organizations running GPT-5.5 or similar models, SkillOpt offers an accessible path to performance improvement. The 23-point gain on procedural tasks translates to