A Developer’s Guide to Pruning Codex Skills for Stronger Models
In a Juejin commentary, the author argues that GPT-5.6, K3 and Claude Fable5 can handle many basic development steps that older Skills packages explicitly encoded. Using Codex as the main example, the author recommends retaining only Skills that add scripts, templates, specialist knowledge, personal preferences or reusable checks, while placing routine project rules in AGENTS.md and mechanical safeguards in Hooks, CI, linters or tests.
The article offers a concrete framework for reducing agent workflow overhead: keep Skills that provide project-specific or hard-to-infer capabilities, and move routine rules and mechanical checks to more appropriate tools.
前几天在知乎看到有人问：Codex 用上 GPT-5.6 后，为什么不少人开始不用 Skills 了？
确实，相信大家也发现，很多 Skill，尤其是开发相关的，真的慢慢就开始用不上了。
我现在看到一个 Skill，会先看它到底能提供什么能力。如果没有确切作用的话，肯定是不会装的。
前两年，模型干活经常漏步骤，Skill 写得越细越让人安心。现在 GPT-5.6 、K3、Claude Fable5 等模型已经很强，能自己补上大部分基础动作，手里的 Skill 也该重新过一遍。
下面我会以 Codex 作为 Coding Agent 为例来谈，其他都是类似的。
#Codex#GPT-5.6#K3#Claude Fable5#Superpowers#mattpocock/skills#AGENTS.md#Claude Code#Cursor#draw.io