Microsoft shows AI skills can move between different models

Microsoft researchers have found a way to share learned skills between different AI models. When they trained a skill using one model, it still worked when they moved it to a completely different system. This means AI agents might not need to relearn tasks from scratch every time.

The most impressive result happened with spreadsheets. A skill built for the Codex model boosted performance on Claude Code from 22 percent to over 81 percent. The transferred skill even outperformed the system that the new model trained for itself.

However, this does not work perfectly for every task yet. While spreadsheet performance was great, the success rate for math dropped significantly. This shows that while transferring skills is possible, it is not a one size fits all solution just yet.