The practical problem with out-of-the-box Claude responses is well-documented among power users: excessive bullet lists, filler phrases like "Certainly!" and "Great question!", and a tendency to pad answers with structure that adds length but not clarity. Claudette (GitHub: adnanakil/nobuzz) is a curated set of system prompts designed to suppress exactly these habits.
The core mechanic is straightforward — you prepend one of the provided system prompts before your actual instructions. The prompts explicitly tell Claude to avoid sycophantic openers, skip unnecessary headers and nested lists when plain prose works better, and cut hedging language that dilutes technical answers. Think of it as a style guide enforced at the model level rather than edited in post.
Why this matters for builders: LLM output quality is only partly a model problem — it's also a prompting problem. When Claude is integrated into pipelines, documentation tools, or coding assistants, verbose and performatively cheerful responses create noise that downstream processing has to filter out. Cleaner output at the source reduces that overhead.
The prompts are composable and short enough to inspect and modify without much effort. If your use case demands a specific tone — say, terse CLI-style answers versus structured technical docs — you can adapt the base prompts rather than starting from scratch. The README includes several variants targeting different failure modes.
For teams already using Claude via API, dropping one of these system prompts into your request wrapper is a five-minute change worth testing. Measure response length and user satisfaction before and after — anecdotal reports from the Hacker News thread (236 points, 165 comments) suggest the difference is noticeable enough that many developers have made it a permanent part of their setup.