<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: kaju</title>
    <description>The latest articles on DEV Community by kaju (@kajuiwa_en).</description>
    <link>https://dev.to/kajuiwa_en</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4257289%2Fa4757aa4-f375-482a-929b-cda057551071.jpg</url>
      <title>DEV Community: kaju</title>
      <link>https://dev.to/kajuiwa_en</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/kajuiwa_en"/>
    <language>en</language>
    <item>
      <title>We Made Claude Code and Codex Fight Over the Same Task — With a Score Sheet</title>
      <dc:creator>kaju</dc:creator>
      <pubDate>Sat, 21 Nov 2026 13:17:41 +0000</pubDate>
      <link>https://dev.to/kajuiwa_en/we-made-claude-code-and-codex-fight-over-the-same-task-with-a-score-sheet-536p</link>
      <guid>https://dev.to/kajuiwa_en/we-made-claude-code-and-codex-fight-over-the-same-task-with-a-score-sheet-536p</guid>
      <description>&lt;p&gt;Every week someone posts "I switched from Claude Code to Codex" (or the reverse), and the evidence is always the same: vibes.&lt;/p&gt;

&lt;p&gt;We run a small AI team — a personal AI secretary system with several agents — and we got tired of vibes. So we started running &lt;strong&gt;same-task battles&lt;/strong&gt;: both agents get the &lt;em&gt;identical&lt;/em&gt; task, the &lt;em&gt;identical&lt;/em&gt; prompt, and the &lt;em&gt;identical&lt;/em&gt; input. Then we score them against an answer key we wrote &lt;strong&gt;before&lt;/strong&gt; the battle started.&lt;/p&gt;

&lt;p&gt;Here is exactly how it works, with a worked example.&lt;/p&gt;

&lt;h2&gt;
  
  
  The four fairness rules
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;One prompt.&lt;/strong&gt; Both agents receive the same instruction text, character for character. No tuning the prompt to each model's strengths.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Golden answer first.&lt;/strong&gt; Before either agent runs, we write the ideal output (or the checklist it must satisfy). If you write the answer key afterwards, you will unconsciously grade toward whoever went first.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Mechanical scoring.&lt;/strong&gt; A 100-point rubric (below). Anything that can't be scored mechanically gets cut from the rubric.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rework counts.&lt;/strong&gt; If an agent's output needs fixing, the fix is part of its score — not a footnote. "It got there eventually" is a cost, and costs belong in the result.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The rubric
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Category&lt;/th&gt;
&lt;th&gt;Points&lt;/th&gt;
&lt;th&gt;What we check&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Accuracy&lt;/td&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;td&gt;Matches the golden answer: facts, figures, required items&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Rework&lt;/td&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;Deductions for every correction a human had to make&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Speed&lt;/td&gt;
&lt;td&gt;15&lt;/td&gt;
&lt;td&gt;Wall-clock time, scored relative to the faster agent&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Quality&lt;/td&gt;
&lt;td&gt;15&lt;/td&gt;
&lt;td&gt;Structure, readability, formatting as specified&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Instruction-following&lt;/td&gt;
&lt;td&gt;10&lt;/td&gt;
&lt;td&gt;Did exactly what the prompt said — no more, no less&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  A worked example (sample data)
&lt;/h2&gt;

&lt;p&gt;Task: &lt;em&gt;"Turn these meeting notes into a decision summary: decisions, owners, deadlines. Nothing else."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;The (fictional) notes mention a launch date of March 3, two owners, one deadline — plus a long digression about office snacks. The golden answer lists exactly the three decisions, both owners, and the deadline.&lt;/p&gt;

&lt;p&gt;What typically happens:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Agent A&lt;/strong&gt; returns a clean summary, correct date, both owners — but adds a "Risks" section nobody asked for. Accuracy 40, instruction-following −4.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Agent B&lt;/strong&gt; returns exactly the requested format, but writes the deadline as March 13. Accuracy −10 for the wrong date, no matter how pretty the formatting is.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Final scores might land at 96 vs 89 — and crucially, &lt;em&gt;you can point at why&lt;/em&gt;. The score sheet survives the argument.&lt;/p&gt;

&lt;h2&gt;
  
  
  Our series so far
&lt;/h2&gt;

&lt;p&gt;Real battles, same rules:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Battle 1 — meeting-notes cleanup (writing task): Claude Code 99, Codex 92.&lt;/strong&gt; Nearly perfect on accuracy; the gap came from rework.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Battle 2 — CSV to monthly aggregation (implementation task): Codex 96, Claude Code 93.&lt;/strong&gt; The rematch flipped the winner. Task type matters more than brand loyalty.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Series record: 1–1. The honest conclusion after two battles is that "which agent is better" is the wrong question. "Which agent is better &lt;em&gt;at this task type, under this rubric&lt;/em&gt;" is one you can actually answer — and route work with.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it yourself
&lt;/h2&gt;

&lt;p&gt;We packaged the whole method — the shared prompt template, the 100-point rubric with the scoring rules for speed and rework, and a how-to for building golden answer keys — as a small kit:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://kajuiwaen.gumroad.com/l/jzzrfbw" rel="noopener noreferrer"&gt;Same-Task Battle Kit on Gumroad&lt;/a&gt;&lt;/strong&gt; ($5)&lt;/p&gt;

&lt;p&gt;It's the exact files we use for the series, not a summary of them. Run one battle with your own task, and you'll know more about your agents than a month of vibes will tell you.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;We write about running a personal AI secretary and its agent team — the illustrated series is on note (English edition: kajuiwa_en) and Medium.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claudecode</category>
      <category>codex</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
