How to stop Claude from saying load-bearing
Large language models like Claude are increasingly criticized for a distinctive, over-engineered writing style full of repeated metaphors and stock phrases (“load-bearing,” “belt-and-suspenders,” “honest take,” etc.). Commenters worry this clichéd “AI slop” is leaking into human communication, degrading documentation and prose, and making it harder to trust or even tolerate text online, while others note that these tics are side effects of RLHF and feedback loops rather than any deep intent. Proposed fixes range from prompt engineering and personal style guides (CLAUDE.md) to regex-based post-processing and external “debabbling” tools, though many argue the real solution must come from changing how models are trained and tuned.
Complaints about Claude’s language style
- Many commenters are fatigued by recurring “Claude-isms”: e.g., load‑bearing, belt‑and‑suspenders, honest take, genuine, smoking gun, blast radius, spine, substrate, ledger, surface, gate, seam, planes, scaffolding, and heavy use of em‑dashes.
- Objections are less about any single phrase and more about repetition, grandiose tone, and vague management‑speak that obscures simple points or overstates certainty.
- Some note inappropriate intensifiers (“incredibly cheap,” “bulletproof,” “root cause”) and invented jargon that feels like puzzles rather than explanations.
- Others are unbothered or even like the jargon, saying it’s concise, expressive, and useful in engineering contexts when not overused.
Workarounds and tools
- People use CLAUDE.md / instructions to ban phrases, avoid first‑person pronouns, or demand concision and “normal” tone. Results are mixed; the model often slips.
- Hooks / post‑processing filters (regex replacements, local scripts, browser tools like debabel, caveman, “humanizer,” custom prose linters) are popular because they don’t affect the reasoning process, only visible output.
- Some worry that too many style rules inside prompts may dilute attention and harm task quality.
Theories on why models speak this way
- Common view: RLHF and preference training reward reassuring, “smart‑sounding” corporate style, creating feedback loops where certain tics (“honest,” “load‑bearing”) get over‑amplified.
- Others tie it to mode collapse and synthetic data: models train on each other’s outputs and reinforce the same patterns.
- A minority suggests the model may be overusing under‑used but semantically handy metaphors like “load‑bearing,” though others push back as anthropomorphizing.
Impact on humans and culture
- Several report coworkers and online writers unconsciously adopting Claude/GPT phrasing; AI style is leaking into human prose in docs, blogs, and commits.
- Some see recognizable LLM style as a helpful “tell” to filter AI content; others resent having their own long‑standing habits (em‑dashes, “load‑bearing”) now flagged as AI‑like.
- There’s strong resentment toward AI‑written internal docs and handoffs: verbose, poorly structured “slop” that wastes time. Some teams consider norms like “don’t ship text you didn’t write.”
- Broader split: one camp wants highly capable but obviously non‑human‑sounding tools; another wants specialized models or modes that prioritize good, varied prose for writing tasks.