A Hacker News thread on stopping Claude from saying "load-bearing" reached 413 points and 472 comments. The discussion centers on a repeatable system prompt that eliminates the phrase without harming output quality.
The Phrase Problem
Claude inserts "load-bearing" in technical explanations at rates reported by multiple users as 4-7 times per 800-word response. The word functions as a filler that signals structural importance but adds no new information.
The pattern appears most in architecture, code review, and system design answers.
How the Fix Works
The solution is a single added instruction placed in the system prompt or at the start of a conversation. It explicitly bans the phrase and supplies two replacement constructions.
Users report the change takes effect immediately on the next generation. No temperature or model version changes are required.
How to Try It
Add this line to your system prompt:
"Never use the word 'load-bearing'. Replace any intended use with either 'critical' or a direct description of the component's role."
Test on a 500-word technical question. Most users see zero occurrences after one try.
Pros and Cons
- Pros: Zero cost, works on Claude 3.5 Sonnet and Opus, preserves response length and structure.
- Cons: Requires adding the line to every new thread; does not transfer to other models automatically.
Early testers note the instruction occasionally makes Claude slightly more direct in lists.
Alternatives and Comparisons
Other models show different filler patterns. GPT-4o favors "crucial" while Gemini 1.5 Pro repeats "foundational."
| Model | Common Filler | Fix Method | Extra Tokens |
|---|---|---|---|
| Claude 3.5 | load-bearing | Phrase ban | 0 |
| GPT-4o | crucial | Style instruction | +3-5 |
| Gemini 1.5 | foundational | Role definition | +2-4 |
The Claude-specific ban remains the shortest intervention among the three.
Who Should Use This
Developers writing technical documentation or code reviews with Claude benefit most. Skip the instruction if your workflow already includes heavy post-editing or if you prefer Claude's current voice.
Bottom Line / Verdict
The single-sentence ban removes the repetitive phrase while leaving every other behavior unchanged.
The same approach can be adapted for other overused terms once they are identified in a model's output distribution.
Top comments (0)