The latest articles on DEV Community by minus1ms (@minus1ms). https://dev.to/minus1ms 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%2F4213982%2F7981a608-0bc9-45ad-9684-54790c2f852f.png https://dev.to/minus1ms en minus1ms Wed, 28 Oct 2026 23:19:08 +0000 https://dev.to/minus1ms/stop-claude-code-from-saying-load-bearing-with-a-tiny-hook-1oc0 https://dev.to/minus1ms/stop-claude-code-from-saying-load-bearing-with-a-tiny-hook-1oc0 <p>If you use Claude Code a lot, you've read "load-bearing" more times than you'd like, along with "You're absolutely right" and "delve". The upstream issue (<a href="https://github.com/anthropics/claude-code/issues/53454" rel="noopener noreferrer">anthropics/claude-code#53454</a>) has 120+ upvotes, and commenters measured it getting more frequent with longer contexts and newer models. Putting "never say load-bearing" in CLAUDE.md doesn't stick.</p> <h2> Why a hook works when instructions don't </h2> <p>Claude Code runs your hooks at fixed points. A <strong>Stop</strong> hook runs when Claude is about to end its turn. If it prints <code>{"decision":"block","reason":"..."}</code>, Claude keeps going and reads the reason as feedback. A <strong>PostToolUse</strong> hook that exits with code 2 sends its stderr back to Claude after a Write or Edit.</p> <p>So you don't need the model to remember a rule. You check its output and send it back when it slips.</p> <h2> Phrasebook </h2> <p><a href="https://github.com/minus1ms-dev/phrasebook" rel="noopener noreferrer">Phrasebook</a> is about 100 lines of Python built on that idea:</p> <ul> <li> <strong>Reply check:</strong> if the final message contains a banned phrase, the stop is blocked once and Claude resends the same reply, reworded, without redoing any work. The <code>stop_hook_active</code> flag means it retries at most once, so it can't loop.</li> <li> <strong>File check:</strong> Write/Edit/MultiEdit content is scanned, and Claude is told which phrase in which file to reword.</li> <li> <strong>Matching</strong> is case-insensitive and whole-word only, treats "load bearing" and "load-bearing" the same, accepts both ' and ’, and allows regexes via lines starting with <code>re:</code>.</li> </ul> <h2> A real run </h2> <div class="highlight js-code-highlight"> <pre class="highlight plaintext"><code>prompt: Reply with exactly: "The cache is the load-bearing part of this design." reply: The cache is central to this design. cost: one extra short response (about $0.002 on Haiku) </code></pre> </div> <h2> Install </h2> <p>Inside Claude Code:<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight shell"><code>/plugin <span class="nb">install </span>phrasebook <span class="nt">--marketplace</span> minus1ms-dev/phrasebook </code></pre> </div> <p>Put your own list in <code>~/.claude/phrasebook.txt</code> (one phrase per line). <code>PHRASEBOOK_DISABLE=1</code> turns it off.</p> <p>The code is MIT on <a href="https://github.com/minus1ms-dev/phrasebook" rel="noopener noreferrer">GitHub</a>. It's free; if it saves your sanity, there's a <a href="https://minus1ms.gumroad.com/l/phrasebook" rel="noopener noreferrer">pay-what-you-want page</a>.</p> claude ai productivity python