The most worth studying in Claude Code is not the model itself, but the skills. I think Skills is similar to the current app in stores. In the future, there will be many Skills, and soon there will be paid stores for paid Skills. After reading the internal experience shared by Anthropic engineer Thariq, I think the most important point is: Skill is not just a prompt word, but a reusable unit of agent work. It can include: Markdown instructions, scripts, templates, reference documents hooks、 Configure files and even persist data. This means that the value of Skill lies not in "teaching the model what to say", but in encapsulating the team's validated methods, processes, constraints, and tools into the model for repeated execution. The most valuable practical experiences shared in this article are: 1. Don't write about common sense, write about "counterintuitive experience" Claude already has a lot of programming knowledge. The most valuable parts of Skill are the pitfalls, boundaries, default agreements, and failure cases that are only known within the team. 2. Gotches is the part with the highest information density in Skill A good skill is often not about a lengthy introduction, but about 'where Claude is most likely to go wrong'. These gotchas should be continuously replenished with usage. 3. Skill needs to use the file system for progressive exposure Skill is a folder, not a single MD. You can open detailed API signatures, templates, scripts, examples, and assets for Claude to read as needed. This is much stronger than stuffing everything into one prompt. 4. The ROI of the validation skill class is extremely high Anthropic even suggests that it is worth having engineers spend a week polishing the verification skill. The reason is simple: there are many agents who write code, but very few agents can stably prove that the code is truly usable. 5. The description is for the model, not for people Claude will scan all Skill descriptions when starting and then decide when to call them. So the focus of the description is not on "summarizing the content", but on "telling the model which scenario should trigger this skill". 6. Skill can have memory For example, the standing post skill can write historical output into a log, and the next time it runs, it will know what has changed compared to yesterday. Anthropic also specifically mentioned that persistent data is best placed in a stable directory rather than the skill pack itself. 7. Giving Claude a script is more effective than giving it a lengthy explanation Script helper functions、query library， Essentially, it is all about reducing Claude's cost of inferring and rebuilding the boiler from scratch each time. This way, its token can be used for judgment and combination, rather than duplicating the wheel. 8. Some hooks are suitable for on-demand activation and not suitable for permanent residency For example,/critical can block dangerous commands, and/freeze can limit the write file range. This type of ability is strong, but keeping it on can affect efficiency, so it is more suitable as an on-demand activation skill hook. Skill needs to be managed, otherwise it will soon be abandoned Anthropic also mentioned that bad skills and repetitive skills can easily accumulate over time. Their approach is not to enforce centralization from the beginning, but to first try it out on a small scale, run real usage, and then decide whether to enter the marketplace.