If you've spent any time reading outputs from Claude, GPT-4, or other large language models, you're probably familiar with what some developers have started calling "Claudish"—that distinctive AI voice characterized by excessive hedging, over-explaining caveats, and a verbosity that prioritizes covering all bases over actually getting to the point. A new GitHub gist published this week attempts to solve this problem by teaching Claude to write more like Paul Graham, the Y Combinator co-founder and essayist famous for his direct, opinionated prose.

What Is Claudish—and Why Does It Matter?

The term "Claudish" has emerged in AI-adjacent communities as a shorthand for the verbose, cautious writing style that many language models default to. This includes phrases like "It's important to note that...," "However, it's worth considering that...," and other constructions that pad out sentences without adding substance. While this hedging behavior makes sense from a risk-aversion perspective—AI companies want to avoid making claims that could be proven wrong—it often produces text that's tedious to read and lacks the confident assertions that make good writing memorable. Paul Graham's essays, by contrast, are known for their clarity, directness, and willingness to make bold claims without excessive qualification.

The GitHub Project

The project, titled "Reduce Claudish: Paul-Graham-writing-style-in-Claude.md," appears on GitHub as a detailed prompt or system instruction designed to override Claude's default writing conventions. Rather than explicitly listing rules about what to avoid, the approach seems to work by providing examples and framing that encourages more natural, essay-like output. The timing is notable—while various "prompt engineering" techniques have circulated since LLMs became mainstream, this one stands out for targeting a specific stylistic problem rather than improving factual accuracy or task completion.

Why This Matters for AI Writing

The broader implications here are significant. As organizations increasingly use LLMs to generate content—from marketing copy to technical documentation—the "Claudish" problem represents a real friction point between what AI models naturally produce and what human readers actually want to consume. If prompts like this gain traction, we might see a shift toward AI-generated text that's less safe but more engaging. That's a trade-off worth watching, especially as the underlying models continue to improve and can hopefully achieve directness without sacrificing accuracy.

Key Takeaways

  • "Claudish" describes the verbose, over-hedged writing style common in LLM outputs
  • Paul Graham's essay style is characterized by directness, bold claims, and minimal hedging
  • This GitHub gist provides a prompt-based approach to changing Claude's output voice
  • The project highlights growing demand for AI text that reads like human writing rather than AI-generated content

The Bottom Line

This isn't just about one developer's pet peeve—it's a signal of where user expectations are heading. People want AI that writes like confident humans, not like cautious lawyers covering their assets. If Anthropic and other AI labs want their models to produce genuinely useful prose, they'll need to reckon with the fact that safety-through-verbosity has real costs.