Claudette: Make Claude stop talking like a BuzzFeed article
Many developers are frustrated with Anthropic’s Claude models adopting an overly verbose, moralizing, “BuzzFeed-style” tone, especially in code comments and technical writing. Tools like “Claudette” and similar wrappers attempt to post-process Claude’s output—often via another LLM—to strip out cloying prose and enforce concise, neutral language, while others rely on elaborate prompts or switch to competing models instead. The exchange highlights a broader tension between safety/engagement-focused model personalities and professionals who want terse, task-focused assistants they can reliably control.
Overall sentiment on Claude’s writing style
- Many find Claude’s “BuzzFeed / thought‑leader” prose verbose, hypey, and mentally exhausting, especially in Opus 5 (and to a lesser extent 4.8).
- Complaints include: overuse of superlatives, “load‑bearing” metaphors, faux‑dramatic chaptering, and long explanations where users want terse answers.
- Several users report obnoxious moralizing or lectures (e.g., on safety, “right and wrong”) that burn tokens instead of doing the requested task.
- A minority say they barely notice, don’t mind, or even find the verbosity helpful for nuance and transparency.
Anthropic’s persona and possible causes
- Claude is seen as deliberately anthropomorphized as an “always positive, helpful engineer” rather than a terse tool.
- Some hypothesize reinforcement learning from social‑media style content and automated “engagement” use cases, leading to performative, story‑like output.
- Others speculate about incentives: saving tokens, obscuring reasoning to hinder distillation, or prioritizing agentic coding over prose quality.
- There is disagreement on how many users truly care; some claim it’s a vocal minority, others point to silent switching to competitors.
Workarounds and tooling
- Multiple tools post‑process Claude’s output with another LLM (Gemini or local models) to strip “Claudish” style (e.g., Claudette, Vomit, claudish‑to‑english, /unslop‑like skills).
- Some instead use system prompts and CLAUDE.md/AGENTS.md rules: enforce short comments and strings, ban temporal/narrative comments, restrict word counts, or demand simple/technical English.
- Many report these style constraints erode as context grows or after compaction; hooks, linters, tests, and post‑edit filters are used to enforce rules deterministically.
Impact on coding workflows
- Claude is praised for coding ability but criticized for:
- verbose, transient, or historical code comments (e.g., narrating refactors, referencing ephemeral docs or tickets);
- leaking internal reasoning into user‑facing copy and tooltips;
- context loss and style drift in long sessions.
- Some revert to older models (Opus 4.6/4.8), use “caveman” or “concise” modes, or offload explanation and documentation to other models.
Alternatives and switching
- Multiple commenters have moved to or experiment with other models for better prose or less moralizing (OpenAI, Gemini, GLM 5.x, DeepSeek, Qwen, Sol, various local models).
- Others stay with Claude for coding but wrap it with translators, local cleanup agents, or strict comment‑style enforcement.