microsoft/skillopt

SkillOpt lets you automatically improve how well an AI agent performs tasks, without changing the underlying AI model itself. Instead of rewriting code or fine-tuning a model, it trains a plain-text "skill document", essentially a set of instructions or context that guides the agent. The end result is a compact file (called best_skill.md) that you attach to your existing AI tool to make it noticeably better at its job. The approach borrows the discipline of traditional machine learning training. You run the agent through tasks, score how it did, and a separate "optimizer" model reviews those results and proposes small edits to the skill document, adding, deleting, or replacing lines. An edit only gets accepted if it actually improves the score on a held-out validation set. You repeat this over multiple rounds (called epochs), with a "learning rate" controlling how much the skill text can change per step. The key idea is that the skill document becomes the thing being optimized, not the model's internal weights. This would appeal to teams building AI-powered products who want better performance from models they can't or don't want to modify. For example, if you're using GPT or Claude in a customer-support agent, SkillOpt can produce a refined instruction file that boosts accuracy by 20+ percentage points across various benchmarks. A newer feature called SkillOpt-Sleep can even run overnight, reviewing past sessions and consolidating lessons learned into improved skills. What's notable is the deployable artifact: a small text file, typically 300 to 2,000 tokens, that works with the unchanged target model. The project reports best-or-tied results across 52 combinations of models, benchmarks, and execution environments, and claims that optimized skills transfer between different models and tools without needing retraining. It supports multiple backends including OpenAI, Claude, Codex, and others, and includes a web-based dashboard for monitoring the training process.