SkillOptTranslation site

What is SkillOpt?

SkillOpt‌ is a text-space optimizer (Text-Space Optimizer) open-sourced by Microsoft (in collaboration with Shanghai Jiao Tong University, Fudan University, and Tongji University)‌, with the following core focus:

‌Treat the Agent’s skill documentation (skill.md) as trainable ”external parameters.” Just like training a neural network, use trajectory-driven editing combined with validation-gated updates to automatically iterate toward the optimal skill documentation (best_skill.md) without modifying any model weights.‌

In a nutshell:‌ No need to change the model—just update the ”instructions”—so the agent gets smarter the more it’s used.‌

It published a preprint on arXiv on May 26, 2026, released its open-source code on May 22, and achieved first or tied-for-first place across all 52 test cases—comprising 6 major benchmarks × 7 models × 3 execution environments—a milestone hailed by the industry as marking ”the transition of agent skill development from a manual workshop to industrial-scale production.”

Key Features of SkillOpt‌

• ‌Automated Iterative Optimization‌：告别手动反复修改 Prompt 的痛苦循环。只需提供带有标准答案的训练集和验证集，SkillOpt 就能自动执行任务、分析失败原因，并精准修改技能文档。
• ‌Strict verification and access control mechanisms‌: Every change made to the skill documentation must pass testing on a separate validation set; the change will only be retained if the score improves; otherwise, it will be rolled back to the previous version. This ensures that the skill will only get better over time, rather than worse.
• ‌Easy Deployment and Visualization‌: Once the training is complete, you will receive a compact best_skill.md When deploying, simply inject this file as a system prompt. It also provides a WebUI panel that allows you to monitor iteration progress, view revision history, and compare results in real time.
• ‌Support for Multiple Models and Backends‌: Supports APIs for various mainstream models, including OpenAI, Anthropic, Qwen, and MiniMax, and works across different execution environments (such as direct conversation, Codex CLI, and Claude Code CLI).

SkillOpt’s Core Technology

At the core of SkillOpt is the integration of deep learning training principles into the text domain, resulting in a closed-loop iterative system with a division of labor between two models.

• ‌Dual-Model Collaboration‌:
  • ‌Execution Model‌: Responsible for performing tasks using the current skill documentation and recording a complete history of the execution and the results.
  • ‌Model Optimization‌: Review the execution history, analyze cases of success and failure, and make precise modifications to the technical documentation using structured add, delete, and update operations.
• ‌A Deep Learning-Style Training Loop‌: The entire optimization process follows a rigorous iterative cycle:‌Rollout (Execute Task) → Reflect (Reflect and Analyze) → Aggregate/Select (Aggregate, Filter, and Modify Recommendations) → Update (Update Skills) → Validate (Verify Results)‌.
• ‌Text Learning Rate and Editing Budget‌: To control the extent of modifications and avoid making too many changes at once that could disrupt the original effective rules, SkillOpt introduces the concept of a “text learning rate” or “editing budget,” which limits the number of modifications to a finite number of steps per iteration.
• ‌Reject Edit Buffer‌: Proposed changes that are rejected during the validation phase are stored in a buffer as negative feedback to prevent future optimizations from repeating the same mistakes.
• ‌Slow/Yuan Update‌: At the end of each iteration, a comprehensive review is conducted to ensure that the skill documentation is logically coherent, concise, and efficient, and to identify longer-term patterns.

Use Cases for SkillOpt

SkillOpt's project URL

• Project website: https://microsoft.github.io/SkillOpt/
• GitHub repository: https://github.com/microsoft/SkillOpt

Comparison of similar products

Compared to traditional manual optimization or existing methods, SkillOpt’s advantages lie in its systematic approach and reliability.

Skills trained using SkillOpt also have excellent‌Mobility‌ Skills that have been optimized for one model or framework often yield significant performance improvements when applied to another model or framework, without the need for retraining.