Claudette: Make Claude stop talking like a BuzzFeed article

What happened

Here is the article:

---

Claudette: Make Claude Stop Talking Like a BuzzFeed Article

A developer named Adnan Akil has released a small open-source tool, hosted on GitHub under the repository adnanakil/nobuzz, that pipes Claude's output through a second model to strip out the dramatic phrasing, numbered revelations, and solemn kickers that characterise how Anthropic's Claude tends to write. The tool drew 131 points and 96 comments on Hacker News, suggesting the irritation it addresses is widespread among engineers who use Claude Code daily.

How it works

This piece looks at what nobuzz does, how it is wired together, and what it means for developers who depend on Claude Code as a primary coding assistant. If you have ever asked Claude why a test is flaky and received a three-part narrative with a "most instructive yet" twist, this is aimed at you.

What happened

Akil built a Claude Code skill — a custom slash command he refers to as /debuzz — that takes Claude's most recent response and sends it to Gemini CLI with a single instruction: say the same thing like a human. The repository is called nobuzz but the tool carries the unofficial name Claudette, which the README notes cannot possibly be its real name because Cat Wu is famously litigious. The project's before-and-after example shows a Claude response describing a "load-bearing assumption" with "three things that jumped out" and a "kicker" being reduced to a plain statement of three concrete bugs in a sync pipeline.

The post earned 131 points on Hacker News item 49388752, and the 96-comment thread shows that the frustration runs well beyond personal preference. Multiple commenters describe the problem showing up in generated code comments, user-facing tooltip strings, and documentation, not just in conversational replies.

Why it matters

How it works

The /debuzz skill is invoked inside a Claude Code session after Claude produces a response. Claude Code passes that response to Gemini CLI, which has been given one job: translate the text from what the README calls millennial clickbait to regular English. The translation is then printed verbatim by Claude Code without further editing, because as the README explains, allowing Claude to "tidy up" the Gemini output reintroduces the voice that was just removed. The constraint that Claude must print the translation verbatim is load-bearing: without it, Claude would smooth the output back into its preferred register.

The architecture is deliberately two-model. One model does the work and writes in its natural style; the other model rewrites the prose. Neither model is instructed "do not be dramatic," which, as a commenter named YuriNiyazov observed in the thread, risks triggering the opposite behaviour by priming the model to think about the exact trait being suppressed. Instead, the second model simply rewrites from scratch with no awareness of the instruction being enforced.

Why it matters

Who is affected

Claude's verbosity is not purely an aesthetic complaint. One commenter with the handle ianjbutler pointed out that externalized chain-of-thought in code comments confuses downstream model runs, wastes tokens, and creates false ground-truth assumptions that propagate through an agentic codebase. A single comment narrating why a choice was made during a specific session becomes stale the moment the session closes, yet it reads as authoritative to any model that processes the file later. The nobuzz approach treats the style problem as a pipeline problem rather than a prompting problem, which is a meaningful shift in how developers think about controlling model output.

The Hacker News thread also referenced a separate project that appeared on HN 23 hours earlier: "Vomit: Clean up Claude 5's token output with a separate LLM" at github.com/zachahn, which scored 285 points and 288 comments. Two independent tools solving the same problem within 24 hours, both relying on a second model to filter the first, suggests that prompting alone has not been sufficient and that the pattern of using one model to police another's prose is becoming an established workaround.

Who is affected

The primary audience is developers using Claude Code — Anthropic's terminal-based agentic coding assistant — as a day-to-day pair programmer. The problem is concentrated there because Claude Code runs long sessions in which the model's explanatory style accumulates across dozens of responses, code comments, and generated documentation strings. A commenter named jorl17 described Claude writing tooltip labels that expose every internal edge case to end users, and user-facing strings with the same verbose register as its developer-facing explanations.

What to watch next

Developers who use Claude through the standard chat interface at shorter session lengths may find the problem less acute, though the thread suggests the style is model-wide and not specific to Claude Code. Teams that review AI-generated code in pull requests are also affected, because the commenting style makes diffs noisier and embeds session-specific reasoning in files that will outlive the session.

What to watch next

The approach nobuzz uses — routing output through a second model for style correction — is still in its early form. A builder trying to replicate or extend it should verify that the Gemini CLI command being called is pinned to a specific model version, since a different Gemini release could produce its own stylistic artefacts. It is also worth checking whether the verbatim-print constraint holds when Claude Code compacts its session context, because the thread shows that style instructions placed in CLAUDE.md files are frequently ignored after compaction events.

Anthropic has not commented on the thread. The broader question the project raises is whether a model that consistently resists style instructions in its own context can be corrected at inference time without a dedicated filtering layer, or whether the correction needs to happen earlier, during training or fine-tuning. Given that "Vomit" and nobuzz appeared independently within a single day and together attracted over 400 Hacker News points and nearly 400 comments, it is a question enough developers are willing to pay two API calls to answer.