SkillOpt is a novel training framework that treats an AI agent's skill documentation (skill.md) like neural network weights, allowing for significant performance improvements without touching the model itself.
Imagine you’ve hired a new employee, and you have them read the same work manual every day so they can perform their tasks well. But every time this employee makes a mistake, you meticulously revise the manual sentence by sentence. Ultimately, the employee’s intelligence remains the same, but as the manual becomes perfect, their work performance improves dramatically. This is the principle behind ‘SkillOpt,’ which we will introduce today.
Recently, the field of artificial intelligence (AI) has been focusing on technologies that improve the performance of AI agents (AI programs that autonomously perform specific objectives) without directly modifying large models. Instead of re-training the AI, these technologies boost its capabilities by intelligently changing the ‘work guidelines’ that the AI reads.
Why is this important?
Until now, improving AI performance required ‘fine-tuning,’ which involves re-training the model itself. However, fine-tuning requires tremendous costs and time, and it often leads to side effects such as changing the model’s existing personality or degradation of specific abilities.
SkillOpt skips this complex process. This is because it focuses exclusively on optimizing the ‘skill document (skill.md),’ which the AI agent references while working Reference 13. This means that companies or developers can build AI specialized for specific tasks quickly and efficiently without replacing the entire model Reference 5.
Understanding it easily
To understand how SkillOpt works, think of this process as a ‘study method learning loop’ for AI Reference 1.
This technology undergoes a four-stage training loop Reference 4:
- Rollout: The agent performs various tasks based on its current skill document. During this process, it records in detail which tools it used and what messages it sent Reference 4.
- Scoring: The agent’s work results are scored. This is the stage where it is confirmed what went well and what went wrong Reference 7.
- Reflection: Based on the scores, another optimization model ponders, “Where could we fix things to do better?” Reference 7. It is like us creating a note for incorrect answers Reference 1.
- Update: Based on those reflections, the content of the skill document is added to, deleted from, or made clearer Reference 7.
To use an easy analogy, the model (AI) is the ‘intelligence,’ and the skill document is the ‘user manual.’ SkillOpt is like having someone rewrite the user manual to be more accurate and effective each time, rather than fixing the AI’s intelligence Reference 10.
Current situation
SkillOpt is already proving substantial performance improvements. Despite never touching the model’s weights, research results show performance gains of 24.8 points for the Codex model and 19.1 points for Claude Code Reference 13.
Currently, AI agents are extending their functionality through a lightweight open format called ‘skill documentation (skill.md),’ and SkillOpt has achieved data-driven optimization automation while following this format Reference 14, Reference 16. The era of ‘text-space optimization’ has opened, where AI learns and improves its own skills without humans needing to manually write instructions one by one Reference 13.
What will happen in the future?
The emergence of SkillOpt will significantly change how AI is developed. Previously, to improve AI performance, prompts (commands) had to be manually tuned like magic spells, but now, performance can be scientifically managed through clear ‘training loops’ Reference 9, Reference 11.
In the future, we will see AI agents create their own optimal work manuals before starting specific tasks, and as they perform the work, the manuals will become increasingly sophisticated. It will become an essential element for AI not only to have ‘intelligence,’ but also a ‘learnable manual’ that helps that intelligence be correctly expressed.
MindTickleBytes AI Reporter’s Perspective
Instead of forcibly opening the black box of complex artificial intelligence technology, SkillOpt has found a smart detour to tame AI through the ‘open door’ of external instructions. This practical approach, which improves results noticeably while lowering technical complexity, will be the key driving force that accelerates the popularization of AI.
References
-
SkillOpt Executive Strategy for Self-Evolving Agent Skills: https://microsoft.github.io/SkillOpt/ -
Microsoft SkillOpt Explained: How to Train AI Agent Skills… Flowtivity: https://flowtivity.ai/blog/microsoft-skillopt-train-ai-agent-skills/ -
Microsoft SkillOpt: The AI Cheat Code for Agent Performance Stork.AI: https://www.stork.ai/blog/microsofts-new-ai-cheat-code -
SkillOpt: Training Loop for Agent Skills AgentBoss: https://agentboss.co/intel/81c44b61be08-skillopt-training-loop-for-agent-skills -
SkillOpt: Training Agent Skill Documents Like… Prompt Genius: https://promptgenius.net/agents/concepts/skillopt -
How to use SkillOpt: train your agent skills, not… The Vibe Company: https://www.thevibecompany.co/en/resources/articles/how-to-use-skillopt-to-train-agent-skills - [2605.23904] SkillOpt: Executive Strategy for Self-Evolving Agent Skills: https://arxiv.org/abs/2605.23904
- GitHub - microsoft/SkillOpt: SkillOpt is a text-space optimizer that…: https://www.linkedin.com/posts/gorebel-artificial-intelligence_github-microsoftskillopt-skillopt-is-activity-7465971443491794944-j8Vt
-
SkillOpt Optimization Techniques to Maximize LLM Agent Performance SuanLab: https://suanlab.com/blog/20260624-paper-2605-23904-skillopt-executive-strategy-fo/ -
Microsoft SkillOpt: The Self-Evolving Agent That Trains explainx.ai: https://explainx.ai/blog/microsoft-skillopt-self-evolving-agent-skills-optimization-2026 - AgentSkills Overview - AgentSkills: https://agentskills.io/
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills - Agentic AI…: https://agentic-ai.readthedocs.io/en/latest/PromptEngineering/skillopt/
- Modifying the weights of the AI model itself
- Iterative learning and optimization of skill documentation (skill.md)
- Manual user prompt entry
- Generates optimal instructions at once without learning
- Humans rewrite all instructions every time
- Iteratively updates instructions by scoring the agent's execution results
- Maintaining the stability of the model architecture
- Making complex fine-tuning processes mandatory
- Rapid increase in hardware costs