nobuzz: Removing 'Claudisms' from Claude Code with Gemini

nobuzz uses a secondary LLM to strip verbose 'claudisms' from technical responses

nobuzz (internally referred to by the author as "Claudette") is a Claude Code skill designed to eliminate the flowery, overly dramatic, and clickbait-style prose often found in Claude's technical explanations. Rather than relying on prompting—which the author argues is insufficient to fully cure the model's tendency toward "TED talk" delivery—nobuzz pipes Claude's output through the Gemini CLI to rewrite the text into plain, human-like English.

The Problem: The "BuzzFeed" Style of Technical Communication

Users of Claude, particularly in engineering contexts, have noted a recurring pattern of verbosity and theatricality. The author describes this as Claude talking like it is "delivering a TED talk about its own pull request," characterized by phrases such as "load-bearing assumption," numbered revelations, and "the kicker."

This style of communication creates cognitive overhead for developers who require concise, factual updates. For example, a standard Claude response might frame a bug as a "load-bearing assumption of the entire sync pipeline," whereas a "debuzzed" response simply lists the bugs, the affected files, and the necessary fixes.

Implementation and Workflow

nobuzz operates as a skill within Claude Code. It does not attempt to change Claude's internal state but instead acts as a post-processing layer:

  1. Capture: The skill writes Claude's previous reply to a temporary file.
  2. Translation: It pipes that file through the Gemini CLI using a specific prompt: gemini -p "<plain-English style instructions>".
  3. Output: Gemini's translation is printed verbatim.

To prevent the "buzzer" from re-introducing the same voice, the skill explicitly avoids letting Claude "tidy up" the Gemini translation.

Available Modes

Depending on the target audience, nobuzz offers three distinct translation modes:

Mode Target Audience Output Characteristics
colleague (default) Engineers Maintains all file paths and code blocks; removes all theatricality.
manager Technical Managers Focuses on what happened, why it matters, and next steps; approximately 1/3 the length of the original.
director Executives 3-5 sentences focusing on outcome, impact, and the "ask."

Technical Requirements and Installation

To use nobuzz, users must have the following installed:

  • Claude Code
  • Gemini CLI (npm install -g @google/gemini-cli), authenticated via /auth or the GEMINI_API_KEY environment variable.

Installation is performed by cloning the repository and copying the skill into the ~/.claude/skills directory.

Community Perspectives on "AI Slop"

The release of nobuzz has sparked a broader discussion among developers regarding the "personality" of frontier models.

The Case for Model Chaining

Some users argue that chaining models is the most practical approach to prompt engineering. One user noted that spending time crafting long essays to tell a model to stop being itself is less effective than simply using a second model to clean up the output.

Alternative Mitigation Strategies

Other developers suggest that system prompts can achieve similar results without the cost of additional tokens. Suggested prompts include:

  • "Use technical language to spell things out, and keep it free of jargon and project shorthand."
  • "For humans output writing at a 10th grade reading level."
  • "Speak in simplified technical English."
  • Limiting word counts (e.g., "Comment blocks are <= 7 words").

Criticisms of the Model's Tone

Several users expressed frustration with the current state of Claude's prose, comparing it to "mental overhead" or "AI slop." Some users reported switching to other models like GPT-4 or GLM5.3 specifically to escape the verbose style, while others noted that this "explanatory tone" has increased in updates around March and July.

Conversely, some users find the "claudisms" helpful, arguing that the specific language used helps them understand the internal meaning the model is assigning to a certain situation.

Related Tools

Similar efforts to "unslop" AI output have appeared in the community, including tools like Vomit (which cleans up Claude 5's output with a separate LLM) and pstack-claude.

Sources

Related

  • Project
    fcakyon/claude-codex-settings

    A collection of plug‑ins that extend Claude Code (and compatible tools) with code‑review, buzzword‑blocking, cross‑model advisors, concise output styles, smarter summarisation, telemetry and office‑document skills.

  • Dispatch
    Removing Claude AI Claudisms: Using MessageDisplay Hooks for Vocabulary Control

    Learn how to use the MessageDisplay hook in Claude to programmatically replace annoying AI-specific phrases (Claudisms) like 'load-bearing' and 'honest take' with custom text.

  • Dispatch
    Vomit: Cleaning Up Claude 5 Token Output with Local LLMs

    Vomit is an open-source Go tool that uses a local LLM to rewrite the verbose and often incomprehensible output of Claude 5 into clear English.

  • Dispatch
    Claude for Apple Foundation Models Integration

    Anthropic has released a Swift package allowing Claude to be used as a server-side language model within Apple's Foundation Models framework, providing a unified API for switching between on-device and cloud-hosted models.

  • Project
    alirezarezvani/claude-skills

    A massive open‑source library (364 skills, 120 commands) that adds domain‑specific expertise to Claude Code, Codex, Gemini CLI, Cursor, and 9 other AI coding agents. Each skill bundles a structured markdown guide, pure‑Python CLI helpers, and reference assets. The repo includes personas, an orchestration protocol, and a built‑in security auditor, making it a turnkey way to turn any LLM‑coding assistant into a full‑stack engineering, product, marketing, compliance, or C‑suite advisor.