<?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: Tom Hitchcock</title>
    <description>The latest articles on DEV Community by Tom Hitchcock (@tom_hitchcock_992cde3cb0e).</description>
    <link>https://dev.to/tom_hitchcock_992cde3cb0e</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%2F4243690%2F3ff941a3-8fe7-4247-a0b5-0000743543af.png</url>
      <title>DEV Community: Tom Hitchcock</title>
      <link>https://dev.to/tom_hitchcock_992cde3cb0e</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/tom_hitchcock_992cde3cb0e"/>
    <language>en</language>
    <item>
      <title>Free beta: the guard rails we built after our AI agent's mistakes (Claude Code hooks)</title>
      <dc:creator>Tom Hitchcock</dc:creator>
      <pubDate>Sat, 14 Nov 2026 00:02:13 +0000</pubDate>
      <link>https://dev.to/tom_hitchcock_992cde3cb0e/free-beta-the-guard-rails-we-built-after-our-ai-agents-mistakes-claude-code-hooks-5f8f</link>
      <guid>https://dev.to/tom_hitchcock_992cde3cb0e/free-beta-the-guard-rails-we-built-after-our-ai-agents-mistakes-claude-code-hooks-5f8f</guid>
      <description>&lt;p&gt;I run a small business in Sydney with an AI agent doing a lot of the day-to-day work: emails, posting, scheduled jobs, the lot.&lt;/p&gt;

&lt;p&gt;It's brilliant. It has also:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;sent the same email to someone twice,&lt;/li&gt;
&lt;li&gt;told me a job was "done" when the job had quietly failed,&lt;/li&gt;
&lt;li&gt;burned through a day's budget in a loop that kept retrying the same thing,&lt;/li&gt;
&lt;li&gt;and very nearly tidied up the wrong folder.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Each time, we added a rule so it couldn't happen again. Those rules are now a small kit, and I'm looking for &lt;strong&gt;3 to 5 people who run Claude Code (or any agent that runs shell commands) to try it free&lt;/strong&gt; and tell me honestly whether it helped.&lt;/p&gt;

&lt;h2&gt;
  
  
  What's in it
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Dangerous-command block&lt;/strong&gt; (PreToolUse hook): stops &lt;code&gt;rm -rf /&lt;/code&gt;, force-push, &lt;code&gt;git reset --hard&lt;/code&gt;, &lt;code&gt;DROP TABLE&lt;/code&gt;, &lt;code&gt;DELETE&lt;/code&gt; with no &lt;code&gt;WHERE&lt;/code&gt;, &lt;code&gt;terraform destroy&lt;/code&gt;, &lt;code&gt;curl | sh&lt;/code&gt;, printing &lt;code&gt;.env&lt;/code&gt; and more (18 rules) &lt;em&gt;before&lt;/em&gt; they run. The agent gets told why.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Budget limiter&lt;/strong&gt; (PreToolUse hook): stops loops and runaway sessions: the same call over and over, too many calls per session, hour or day.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;"Done means proven"&lt;/strong&gt; (Stop hook): when the agent says it's finished, your own tests or build run first. If they fail, it goes back to fix them instead of telling you it's done.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Send guard&lt;/strong&gt; (Python): one check before any email or message goes out. It holds duplicates, bursts from a loop, &lt;code&gt;[Name]&lt;/code&gt; left in a template, internal notes, do-not-contact addresses, and 2am sends.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Plain Python, no dependencies, one installer that keeps your existing &lt;code&gt;.claude/settings.json&lt;/code&gt; (and backs it up). &lt;code&gt;python install.py /your/project&lt;/code&gt;, restart Claude Code, done. &lt;code&gt;--remove&lt;/code&gt; takes it out cleanly. It ships with 59 self-tests.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Download (free beta, zip, 15 KB):&lt;/strong&gt; &lt;a href="https://drive.google.com/uc?export=download&amp;amp;id=1_TOoL77LaMYEk02UfQKw9bCvwDlUiBvH" rel="noopener noreferrer"&gt;agent-safety-kit-beta-0.1.zip&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'm asking
&lt;/h2&gt;

&lt;p&gt;Install it on a real project for a week, then comment below:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Did it block anything real? (the command is in &lt;code&gt;.agent-safety/blocked.jsonl&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;Did it block anything it shouldn't have?&lt;/li&gt;
&lt;li&gt;What rule is missing?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Honest limits: pattern rules catch the common destructive commands, not every possible one. It's a seatbelt, not a substitute for backups and least-privilege credentials. Tested on Windows 11 with Python 3.11 so far, so Mac and Linux reports are especially welcome.&lt;/p&gt;

&lt;p&gt;Thanks in advance. Every block it makes on your machine is one less "how did that happen" for you.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>devops</category>
      <category>python</category>
    </item>
  </channel>
</rss>
