I'm done with LLM-through-chat-experience
The author argues that using LLMs through chat interfaces is fundamentally flawed, as it confuses an interface with the technology itself and imposes unnatural conversational patterns on tasks better suited for direct, purpose-built tools. They advocate for moving beyond chat-based interactions to more structured and efficient ways of leveraging language models.
Background
The author argues that interacting with LLMs (large language models like ChatGPT, Claude, Gemini) primarily through chat interfaces is a dead end. Instead of fixing the core friction — having to prompt, explain context, and verify output every time — users are wasting time on "vibes-based" conversation. They advocate for structured, programmatic interaction: giving LLMs clear goals, constraints, and feedback loops without the chat window. Notable is the reference to "Deepak's piece" and Simon Willison's "give it a sandbox and a goal" framing, which are shorthand for a growing critique among AI practitioners that current chat-based UX hides the real power of LLMs as reasoning engines.