How to Stop Claude from Saying 'Load-Bearing'

If you have spent any time with Claude Code, Codex, or really any frontier LLM doing code work, you have noticed the verbal tics. "Load-bearing." "Honest take." "Belt and suspenders." "That's the unlock." "Smoking gun." The phrases show up so often that they start to feel like a shared affliction among developers using AI coding tools.

A blog post from jola.dev hit the Hacker News front page today with a surprisingly practical solution: a Python hook that intercepts Claude's output and swaps the annoying phrases before they hit your terminal. The discussion that followed became a fascinating window into how developers are adapting to (and coping with) LLM-generated prose.

The Technical Fix#

The approach is simple but effective. You create a MessageDisplay hook - a Python script that processes Claude's output through a regex-based text replacement filter. The script lives at ~/.claude/hooks/wordswap.sh, and after configuring it in ~/.claude/settings.json, every response gets filtered before display.

The author's example replacements lean into absurdity: "load-bearing" becomes "cooked," "seam" becomes "whatchamacallit," and "you're absolutely right" transforms into "I'm a complete clown." The humor is intentional - if you cannot escape the slop, you might as well make it ridiculous enough to laugh at.

But the real value is in the technique. You can configure your own replacements - swapping "robust" for "solid," stripping "clearly" and "obviously" entirely, or normalizing "belt and suspenders" to "belt and braces" if you are British and find the American version grating.

What Hacker News Is Saying#

The discussion thread exploded with 229 comments, and the conversation went far beyond the original hook idea.

The "infohazard" theory got traction. One commenter described AI speech as an "information hazard" - the more you read LLM output, the more it infects your own writing and thinking. They wrote:

"I read orders of magnitude more AI-speak - I call it 'babble', or perhaps 'Babel' - than human-written text. I can feel its genuinely honest points, clearly stated, slipping their banal tendrils into my thoughts and inner monologue."

The suggested remedy: deliberately read prose "far from slop" to inoculate yourself, and write manually to force different synthesis patterns.

The RLHF blame game. Multiple commenters pointed fingers at reinforcement learning from human feedback. One noted: "Nowadays, with the focus on agentic use and coding, it seems models have all been RLHF'd to death." The irony is that if nobody likes this writing style, how can it be the result of human feedback? The responses speculated: people like each instance well enough in isolation, but the cumulative effect becomes exhausting. The model learns that these phrases get positive signals, then overuses them.

The dead internet theory made an appearance. One commenter suggested the models are reflecting feedback from other LLMs and bots rather than real humans - "Maybe it's the dead internet. All the bots and other LLMs providing feedback, so in reality it's reflecting the reality in a sense."

Some pushed back on hiding the problem. A contingent argued that making LLM output sound more human is undesirable - they want the ability to identify machine-generated text. "I don't want LLMs sounding human. I want the ability to shame and discredit anyone passing the job of prose to a machine. There's an art to writing, and hopefully LLMs never truly get it right."

The "just write it yourself" faction. When one commenter asked "What do people do for writing?", another simply replied: "I use a keyboard, personally." Multiple commenters reported that their companies are considering policies along the lines of "Why should I bother to read something you didn't bother to write?"

British vs American English tensions. The "belt and suspenders" vs "belt and braces" divide generated its own subthread, with someone noting that "suspenders" in British English means what Americans call a garter belt - hence the phrase sounding "particularly odd over here."

The Deeper Problem#

The conversation surfaced a tension that goes beyond word choice. As one commenter put it: "LLMs are pattern-extenders that have nothing to say. The training overfitted to the grace notes in good writing. And since LLMs can't wield language with purpose or experience the feeling of the words, they use these devices arbitrarily."

This maps to a real observation about AI coding agents. They see every problem as calling for a "smoke test" or an unnecessary design pattern. The verbal tics in prose are the same phenomenon as architectural over-engineering in code - the model learned that these patterns correlate with approval, so it deploys them indiscriminately.

The proposed solutions ranged from practical (the hook approach) to philosophical (read more Orwell, write more yourself) to resigned (just accept that this is what AI output sounds like and move on).

Why This Matters for Developers#

If you are using Claude Code, Codex, or any AI coding assistant for significant portions of your day, the vocabulary contamination is real. PR descriptions start to sound the same. Documentation reads like it came from a template. Commit messages develop a suspicious uniformity.

The hook approach is a band-aid, but it is a useful one. More importantly, the discussion highlights the need to maintain your own voice when working heavily with AI tools. Read human-written technical writing. Write your own prose sometimes. Notice when "load-bearing" starts creeping into your vocabulary.

Or, as one commenter suggested, just replace it with "cooked" and laugh every time Claude tells you about the cooked authentication layer in your codebase.

Sources#

• Original article: How to stop Claude from saying load-bearing
• Hacker News discussion (229 comments)
• Politics and the English Language - George Orwell (referenced in thread)