On Monday, a developer who goes by the handle jola published a blog post with a disarmingly practical title: “How to stop Claude from saying load-bearing.” The fix was not a prompt-engineering trick, a plea to Anthropic’s product team, or a demand for regulatory intervention. It was a shell script—a few lines of bash that intercept Claude Code’s output and swap the model’s most grating verbal tics for absurdist alternatives. “You’re absolutely right” becomes “I’m a complete clown.” The script, shared under a permissive license, hit 435 points on Hacker News within hours.

The post struck a nerve that has been raw for months. An issue on Anthropic’s own GitHub repository, opened April 26, documents the same frustration: Claude Code cannot stop describing everything as “load-bearing.” The thread has accumulated hundreds of reactions, and as of this week, no official fix has shipped.

But the real story here is not that an AI model has verbal tics. Every large language model does. The story is who fixed it—and what that says about where the control surface of AI products is migrating.

A Shell Script, Not a System Prompt

Anthropic publishes its system prompts. The company has made transparency a brand value, and the diff between Claude Opus 4.6 and 4.7—analyzed in detail by Simon Willison in April—shows exactly how the sausage is made. The prompts are long, carefully worded, and full of instructions about tone, formatting, and helpfulness. They are also, apparently, insufficient to stop the model from reaching for “load-bearing” every time it encounters a dependency.

Why? Because the tics are not in the system prompt. They are baked into the model’s training distribution—an artifact of the fine-tuning data, the RLHF feedback, the thousands of human preferences that taught Claude that “load-bearing” sounds technically competent and “you’re absolutely right” sounds agreeable. You cannot prompt your way out of a statistical regularity that deep. Retraining is expensive and slow. The model’s quirks are, in a very literal sense, load-bearing.

So the fix came from outside. jola’s script uses Claude Code’s own hooks system—a documented but underused feature that lets users run arbitrary commands on the model’s output before it reaches the terminal. The hook is not a hack; it is infrastructure. Anthropic built the pipe. A user decided what should flow through it.

The Voice Layer Moves to the Edge

This is not an isolated incident. It is a pattern that has been gathering force for two years, and it points to a structural shift in how AI products will be built.

For the first generation of consumer AI—ChatGPT, Claude.ai, Gemini—the voice of the model was part of the product. OpenAI spent enormous effort tuning GPT-4 to sound helpful, cautious, and slightly anodyne. Anthropic did the same, with a different flavor: thoughtful, collaborative, prone to describing things as “honest takes.” The voice was the brand. Changing it required changing the model.

But the hooks system, and the broader ecosystem of output-filtering tools that has grown up around it, suggests a different model. The AI company provides the raw capability—the reasoning, the code generation, the analysis. The user provides the voice. Want Claude to sound like a grizzled sysadmin? Write a wordswap script. Want it to sound like a terse copy editor? Same thing. The intelligence is the platform; the personality is a user-space concern.

This is how every mature software ecosystem evolves. Linux gives you a kernel; you pick your desktop environment. Databases give you a query engine; you pick your ORM. The fact that AI voice-tuning is moving to the edge is not a failure of the central product—it is a sign that the product is becoming infrastructure.

Let the Model Be a Model

There is a lesson here for Anthropic, and for every other AI lab still trying to fine-tune the perfect assistant personality. Stop.

The verbal tics that drive users to write shell scripts are not accidents. They are the exhaust of a process that tries to make a general-purpose reasoning engine sound like a specific kind of person—earnest, collaborative, slightly deferential, fond of the word “fundamentally.” That process is doomed to annoy someone, because no single voice pleases everyone. The more you tune for one audience, the more you alienate another.

The market is already routing around the problem. jola’s script is not the first output filter, and it will not be the last. Within a year, the most popular Claude Code extensions will not be for new capabilities—they will be for voice packs. Imagine a Claude that talks like a battle-hardened incident responder, or a character from a novel, or simply a version of itself that has never once uttered the word “seams.” The raw intelligence is the same. The wrapper is yours.

Anthropic should take the hint. Make the default output more neutral. Make the hooks more powerful. Document the output schema. And then get out of the way. The voice layer belongs to the user now. That is not a bug to be fixed. It is the product working as it should.

As one developer posted in a Slack channel for Claude Code power users this week, alongside a screenshot of a prompt template that began with “DO NOT use the following words: load-bearing, honest take, absolutely right, seams, deeply, fundamentally”: “It works about half the time. The other half, I run the script.”

Sources