Read time: ~6 minutes. Key facts:

• Claude has recognizable filler phrases — “load-bearing,” “you’re absolutely right,” “honest take,” “seam(s)” — that recur across outputs.
• Telling it “don’t say X” often backfires: naming a phrase makes it more salient, and negative instructions are a weak steer.
• The reliable levers: few-shot examples of the voice you want (Anthropic calls examples one of the most reliable ways to steer tone), and Claude Code output styles / CLAUDE.md, which append your style to the system prompt every session.
• Last resort: a Claude Code post-processing hook that find-and-replaces the phrases on display — cosmetic, but it works.

Sourcing note: the output-styles and system-prompt mechanics are from the official Claude Code output styles docs and Anthropic’s prompting best practices; the specific filler phrases and the display-hook approach are from jola.dev’s write-up. “Reliable” here means more reliable than a blocklist — no prompt fully eliminates a tic. Links at the bottom.

If you use Claude for writing, you’ve seen it: everything becomes “load-bearing,” every correction earns a “you’re absolutely right,” and the codebase is full of “seams.” It’s a real, widely-noticed pattern. The instinct is to add “stop saying load-bearing” to your prompt — and it mostly doesn’t work. Here’s why, and what actually moves Claude’s writing style.

---

1. The tics, and why “don’t say X” backfires

The commonly-cited offenders (per jola.dev’s write-up and general community observation): “load-bearing,” “you’re absolutely right,” “honest take,” “seam”/“seams.” They show up because they’re patterns reinforced during training — Claude reaches for them the way a person leans on a favorite phrase.

The trap is the obvious fix. Adding “do not use the word load-bearing” to your prompt tends to underperform for two reasons:

• Naming makes it salient. Putting “load-bearing” in the instructions keeps the token front-and-center in context — you’ve drawn attention to the exact thing you want suppressed. Models frequently rebound toward a forbidden word, not away from it.
• Negative instructions are a weak steer in general. “Don’t do X” gives the model nothing to do instead. It’s far easier to follow “write in this voice” (with a target) than “avoid these ten words” (a minefield with no destination).

So the move isn’t a bigger blocklist. It’s giving Claude a positive target for its voice — and putting that target somewhere it’s read every turn.

---

2. The most reliable lever: show, don’t forbid (few-shot)

Anthropic’s own prompting guidance is blunt about this: examples are one of the most reliable ways to steer output format, tone, and structure. A few well-crafted examples (few-shot / multishot prompting) beat paragraphs of rules.

Instead of “don’t say load-bearing,” give Claude two or three short before/after examples of the register you want:

Match this voice — direct, concrete, no filler:

BAD:  "That test is load-bearing, so you're absolutely right to be
       careful about the seam between the modules."
GOOD: "That test guards the module boundary. Changing it risks the
       integration path, so update it deliberately."

Write in the GOOD style.

You’re not blocking words — you’re demonstrating the destination. This is the single highest-leverage change, and it works in the API, in claude.ai, and in Claude Code.

---

3. Make it stick: Claude Code output styles

If you’re in Claude Code, don’t paste your style into every message — bake it into an output style, which the docs describe as modifying the system prompt to set role, tone, and output format. Output styles “change how Claude responds, not what Claude knows,” and they trigger periodic reminders to keep adhering during the session.

Create a custom one as a Markdown file in .claude/output-styles/ (project) with frontmatter plus your instructions:

---
name: Plain
description: Direct, concrete prose with no filler phrases
keep-coding-instructions: true
---

Write in a plain, direct voice. Prefer concrete nouns and verbs.
Favor short declarative sentences.

Here is the target register (match it):
- Instead of "X is load-bearing," write "X guards / controls / is required by ..."
- Instead of "you're absolutely right," just make the correction and move on.
- State disagreement or uncertainty plainly, without hedging preambles.

Set keep-coding-instructions: true so Claude keeps its software-engineering behavior and only layers your voice on top. Because output styles are part of the system prompt (read once at session start), switch or edit it, then /clear or start a new session for it to take effect. Anthropic ships built-ins too (Proactive, Explanatory, Learning), but for tone you’ll want your own.

For project-wide conventions that aren’t purely style, CLAUDE.md is the companion — it adds project context on top of Claude’s default instructions every session. A short “voice” note there reinforces the output style. (Just keep it tight — remember CLAUDE.md is a constant per-turn cost; see reducing Claude Code token overhead.)

---

4. Frame positively in the system prompt (API)

If you’re calling the API directly, the same principle applies in your system prompt: describe the voice, don’t enumerate a ban list. Compare:

• ❌ “Never use the words: load-bearing, honest take, seam, delve, you’re absolutely right…”
• ✅ “Write like a senior engineer explaining to a peer: concrete, unadorned, no throat-clearing preambles, no flattery. Correct mistakes directly.”

The positive version gives a target and doesn’t spotlight the exact tokens you’re trying to avoid. Pair it with the §2 examples for the strongest effect.

---

5. The nuclear option: a post-processing hook

Sometimes you just want the word gone, regardless of what the model generates. That’s what jola.dev did: a Claude Code hook (a shell script at ~/.claude/hooks/wordswap.sh) that intercepts Claude’s output and find-and-replaces the offending phrases before they’re displayed — swapping “load-bearing” for whatever you like (their example: “spicy doodad”).

Be clear-eyed about what this is: it’s cosmetic. It doesn’t change what Claude generates or reasons about — it edits the text on the way to your screen. That makes it a fine last resort for a personal tic you can’t stand, but a bad fit when the wording matters (it will happily mangle a legitimate use of “seam” in a sewing app). Use steering (§2–4) for real behavior change; use the hook only as a display-layer backstop.

---

6. What actually works, in order

1. Give 2–3 before/after examples of your target voice — the highest-leverage move (§2).
2. Bake it into a Claude Code output style (or a positive system prompt via the API) so it’s applied every session, not re-typed (§3–4).
3. Reinforce project conventions in CLAUDE.md, kept short.
4. Only then, if a specific word still grates, add a display hook to rewrite it cosmetically (§5).

Notably absent: a big “banned words” list. It’s the most common attempt and the least effective — it spotlights the words and gives Claude no target to aim at instead.

The takeaway

Claude’s filler phrases — “load-bearing,” “you’re absolutely right,” “honest take,” “seam” — resist the obvious fix, because “don’t say X” makes X more salient and gives no alternative. Steer with a positive target instead: few-shot examples of the voice you want (Anthropic’s most-reliable lever), baked into a Claude Code output style or a positive API system prompt so it applies every session, reinforced by a tight CLAUDE.md. Keep the find-and-replace hook as a purely cosmetic last resort. Show Claude the register you want; don’t hand it a blocklist.

For more on getting the most out of Claude Code, see cutting its token overhead, using it as a daily driver, and why long agent sessions drift.

Sources

• Output styles — Claude Code docs — output styles modify the system prompt to set role/tone/format (“change how Claude responds, not what it knows”); custom styles are Markdown files in .claude/output-styles/ with name / description / keep-coding-instructions frontmatter; part of the system prompt (read at session start; /clear or new session to apply); built-ins Proactive / Explanatory / Learning
• Prompting best practices — Anthropic — examples (few-shot / multishot) are one of the most reliable ways to steer output tone, format, and structure
• How to stop Claude from saying load-bearing — jola.dev — the specific filler phrases (“load-bearing,” “you’re absolutely right,” “honest take,” “seam”) and the ~/.claude/hooks/wordswap.sh display-hook find-and-replace approach
• Steering reduces a tic; it doesn’t guarantee elimination. The hook approach is cosmetic (display-only). Verified July 15, 2026 — confirm output-style frontmatter fields against the live Claude Code docs.