<?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>Headroom on AI Pulse</title>
    <link>https://cuigh.com/tags/headroom/</link>
    <description>Recent content in Headroom on AI Pulse</description>
    <generator>Hugo</generator>
    <language>en-US</language>
    <lastBuildDate>Mon, 22 Jun 2026 10:51:00 +0800</lastBuildDate>
    <atom:link href="https://cuigh.com/tags/headroom/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>90% of my LLM tokens were burning the wrong fires: how headroom cut my agent costs 7x</title>
      <link>https://cuigh.com/posts/headroom-token-90-percent-burned-2026/</link>
      <pubDate>Mon, 22 Jun 2026 10:51:00 +0800</pubDate>
      <guid>https://cuigh.com/posts/headroom-token-90-percent-burned-2026/</guid>
      <description>headroom is the GitHub monthly #1 trending repo (&#43;41,093 stars this month), promising 60–95% fewer tokens before they reach the LLM. I ran a real-world test against one week of my own Claude Code logs. Result: 7.3x cost drop, half the latency, no measurable answer-quality loss. This isn&amp;#39;t a new-model story. It&amp;#39;s the first time context engineering has stepped to center stage.</description>
    </item>
  </channel>
</rss>
