<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Skill-Routing on Xueping's Blog</title><link>https://hellogxp.github.io/tags/skill-routing/</link><description>Recent content in Skill-Routing on Xueping's Blog</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Thu, 25 Jun 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://hellogxp.github.io/tags/skill-routing/index.xml" rel="self" type="application/rss+xml"/><item><title>Compositional Skill Routing: Decomposition Is the Bottleneck</title><link>https://hellogxp.github.io/posts/skill-routing/</link><pubDate>Thu, 25 Jun 2026 00:00:00 +0000</pubDate><guid>https://hellogxp.github.io/posts/skill-routing/</guid><description>When LLM agents need to chain multiple skills, the hard part isn&amp;rsquo;t retrieval—it&amp;rsquo;s decomposition. Oracle experiments reveal a 65-percentage-point gap. SAD (Skill-Aware Decomposition) closes it by correcting granularity, not by boosting model capacity. Context window drops 99.9%.</description></item></channel></rss>