Removing Claude AI Claudisms: Using MessageDisplay Hooks for Vocabulary Control

Users of Claude AI have noted a recurring pattern of repetitive, predictable vocabulary—often referred to as "Claudisms"—that can make AI-generated text feel robotic or unnatural. Common examples include phrases like "load-bearing seams," "honest takes," and "substrate."

To combat this, a programmatic approach using the MessageDisplay hook allows users to intercept and modify Claude's output before it is displayed in the interface.

Implementing a Word-Swap Hook for Claude

To remove or replace specific phrases in Claude's output, you can create a custom script that acts as a filter. This process involves creating a replacement script and registering it in the Claude settings.

Step 1: Create the Replacement Script

Create a file at ~/.claude/hooks/wordswap.sh and make it executable using chmod +x ~/.claude/hooks/wordswap.sh. The following Python script defines a dictionary of phrases to be replaced and processes the incoming JSON data from Claude:

#!/usr/bin/env python3
import json, re, sys

replacements = {
    "seam": "whatchamacallit",
    "you're absolutely right": "I'm a complete clown",
    "honest take": "spicy doodad",
    "load-bearing": "cooked"
}

data = json.load(sys.stdin)
text = data.get("delta") or ""

for phrase, replacement in replacements.items():
    pattern = r"\b" + re.escape(phrase) + r"\b"
    text = re.sub(pattern, replacement, text, flags=re.IGNORECASE)

print(json.dumps({
    "hookSpecificOutput": {
        "hookEventName": "MessageDisplay",
        "displayContent": text,
    }
}))

Step 2: Configure the Settings

Add the script to the hooks block in your ~/.claude/settings.json file to enable the interceptor:

{
  "hooks": {
    "MessageDisplay": [
      {
        "hooks": [
          { "type": "command", "command": "$HOME/.claude/hooks/wordswap.sh" }
        ]
      }
    ]
  }
}

Once configured, start a new session for the changes to take effect.

Understanding "Claudisms" and AI Vocabulary Bias

Predictable speech patterns in LLMs are a result of the model's training and RLHF (Reinforcement Learning from Human Feedback) processes. While humans also have speaking styles, the scale of LLM output—billions of tokens per day—makes these biases highly visible.

Common Claudisms Identified by Users

Users have identified a variety of repetitive phrases that signal AI generation:

• Technical Terms: "load-bearing," "substrate," "wire/wiring," "projection," "frontier," "quiescence."

• Conversational Fillers: "honest take," "genuine result," "honest assessment," "steelman," "production ready."

• Phrasal Patterns: "This will only bite if...", "stress-testing," "matrix," "anchor," "flagging."

The Impact of AI Prose on Human Writing

There is a growing concern that these patterns are leaking into human communication. Some users report seeing these terms in corporate all-hands messages, commit messages, and professional emails, suggesting that LLM vocabulary is beginning to influence human professional discourse.

Alternative Strategies for Controlling AI Voice

Beyond programmatic hooks, users employ several other methods to reduce AI-isms in their output:

• CLAUDE.md Instructions: Adding a "ban list" of phrases to a global configuration file to instruct the model to avoid specific words.
• Iterative Review: Using a sub-agent or a separate "humanizer" skill to review and rewrite the output for a more natural tone.
• Translation Loops: Translating text into another language (e.g., Spanish or French) and then back to English to strip away standard AI phrasing.
• Custom Personas: Instructing the model to refer to itself by a specific name (e.g., "Clod") to avoid the first-person pronouns that often signal AI writing.

Trade-offs and Performance Considerations

Some users argue that influencing the model's communication style may consume valuable context window space or dilute the model's attention, potentially leading to lower quality technical results. There is also a theory that some idioms may carry specific subtext meaningful to the LLM's internal reasoning, and impeding those choices could subtly hinder performance.