Skill Up: how AI builds the skills other agents use

By Turing Post · October 7, 2026.

This episode of Turing Post's "AI Builds AI" series starts from a simple question: what should a coding agent leave behind when it finishes a job? The obvious answer is working code and evidence that it works, but the session can also surface details that are useful for the next task. Hence the idea of asking an agent to examine its own experience and propose improvements to its working environment, which is exactly what Matt Pocock's /retro skill does (published in his GitHub repository). One of that skill's most sensible instructions is to turn mechanically detectable errors into automated checks, and to reserve written guidance only for decisions that require judgment. That example was the trigger for Turing Post to explore the new wave of interest around agent "skills".

The article explains that turning an agent's experience into a procedure, a test, a tool or an environment change requires engineering judgment, and that several research teams are trying to automate that decision. A number of specific projects are cited: Microsoft's SkillOpt, SkillAdam from Renmin University and Tencent, OpenHands' SkillRefiner, Alibaba's skill-up, Sakana's ShinkaEvolve, and Anthropic's and OpenAI's own approaches to skills. The episode promises to cover, among other things: what an agent skill actually contains; why readable instructions can still be hard to use; how SkillOpt and SkillAdam improve a procedure; what SkillRefiner can learn without having to repeat the original work; how to tell whether a skill generalizes; when a skill should become part of the code or of the model's weights; and a development workflow readers can apply to their own agents.

On the technical basis: a skill can be as simple as instructions in a SKILL.md file, accompanied by scripts, examples and references. Its description helps an agent decide when to load it, and the procedure provides the knowledge about how to carry out a specific type of work. The newsletter includes a minimal SKILL.md example (with the name and description fields, and a series of steps for explaining code: state the purpose in one sentence, describe inputs and outputs, and walk through a small example).

As a use case, it mentions an agent reviewing changes to a machine learning pipeline: the skill could include where the evaluation data lives, which checks detect data leakage, how to compare results, and what evidence should be included in the review. The idea is that the model already understands Python and experimental design in general, but needs to discover how those things are implemented in that particular repository.

The article picks up Anthropic's distinction between skills that fill gaps in a model's capabilities and skills that encode a team's preferred process. According to the text, March's skill-creator update added evaluation and benchmarking support, including success rates, elapsed time and token usage, which lets authors check whether a skill helps, whether an edit broke it, or whether a newer model still needs it.

The definition Turing Post proposes is that a skill is not a particularly elaborate prompt, but a reusable procedure that makes a claim about what will remain true in future tasks: the required principles and formats generalize across tasks, while specific file paths and commands belong strictly to the repository. It also connects to Turing Post's MCP guide: an MCP connection can expose experiment logs or a repository's tools, and the skill can explain how to use those capabilities for a specific task; access, procedure and evidence of success are, according to the article, separate design concerns. The newsletter excerpt closes by raising the discovery problem: once the procedure is written, how will the agent know that it is the right skill for the task?

The newsletter ends with a table summarizing the different research approaches mentioned: "learning from new runs" (run tasks, examine failures, edit the skill and test again, with examples such as SkillOpt and SkillAdam), "learning from past work" (analyze recorded runs and results to propose better instructions, with examples such as SkillRefiner and CASD), "learning from existing code" (turn a repository's knowledge into procedures other agents can use, with the Repo-To-Skill example), "checking whether skills help" (compare performance with and without the skill, including on unseen tasks, with Alibaba's skill-up as the example) and "training the procedure into the model" (use the skill's guidance during training and gradually withdraw it, with the SKILL0 example).

Besides the main article, the newsletter includes a promotion for a free, virtual event organized together with O'Reilly, the "Expert Showcase: Agent Memory", which will take place on October 13 at 11 a.m. ET / 8 a.m. PT and will be moderated by Ksenia Se (founder of Turing Post). Participants will include Pete Johnson, Lewis Liu, Shawn Shen, Jason Davenport, Jess Weng, Kris Murphy, Xavier Bull and Sally O'Malley, with perspectives from MongoDB, Microsoft, Memories.ai, Google Cloud, LangChain, NVIDIA, Adarga and Red Hat/OpenClaw. The announced topics include how agents remember without holding on to all the context, how reflection can turn past experience into useful memories, how to measure whether an agent is actually improving, visual memory for physical agents, memory in high-stakes environments, and the trust boundaries that determine what an agent should retain or forget.