Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Large Skill Catalogs Test Claude Code and Codex Context Budgets

AI summary

掘金 · 人工智能 reports that Claude Code and Codex CLI apply different context-budget rules to skill listings as catalog size grows. In a small test using 300 unrelated skills, Claude Code triggered the target skill in one of three runs, while Codex triggered it in all three by shortening descriptions that retained the target keywords; the tested versions were Claude Code 2.1.323 and Codex CLI 0.162.1. Community and publisher observations also showed that many listed skills could lose their descriptions, though the 30,000-character figure was observational rather than an official limit.

In-depth insight

The comparison shows that large skill catalogs can reduce tool-selection reliability and consume substantial context, while making description wording, catalog cleanup and budget settings practical considerations for AI coding workflows.

给 AI 编程工具装 skill(技能包)已经成为不少开发者的日常习惯。所谓 skill,就是一个包含说明文档的文件夹,里面写清楚了「什么时候用、怎么做」。工具每次开工时,不会把所有说明全文塞进上下文,而是先把所有 skill 的「名字 + 一句描述」列成清单交给模型;模型靠描述判断该用哪个,选中以后才去读取完整说明。

听起来很节省上下文,但如果 skill 装得太多,会发生什么?

社区观察:283 个 Skill 里 220 个没了描述

2026-11-12,开发者 Murat Sert 发文提到一个现象:他装了 283 个 skill,结果发给模型的清单里有 220 个只剩名字、没有描述,清单总长度一直卡在约 30,000 字符(他注明「约 30,000 字符」是观察值,并非官方规格)。他还统计到,用户还没说话,首条请求就已经消耗了 74,500 token,其中约 22,000 来自他装的插件、skill、MCP 和 hooks;而在另一个装了 672 个 skill 的旧会话里,有 642 个没有描述。他在文章里表示,没见过官方文档提这件事。

#Claude Code#Codex CLI

Original · 7 min readRead the original
Group chat

No comments yet. Start the conversation.