The AI models inside your coding assistant got dramatically better over the past year. The interface you use to talk to them? It barely moved.
That's the core argument from developer David Reis, who makes a sharp case that tools like Cursor, Claude Code, and their competitors are repeating a familiar pattern: powerful engines trapped behind lazy interfaces. The chat box you type prompts into looks almost identical to what shipped with the first coding assistants, and Reis argues that design choice actively steers developers into workflows that fail.
The Blank Canvas Trap
Open any AI coding tool and you get an empty text box. No structure, no guidance, no guardrails. The natural response is to dump everything into one massive prompt and hope for the best. When the output is wrong, the natural next step is to follow up with corrections, which inflates the context (the running memory of your conversation) until the model starts producing worse results, not better ones.
Reis calls this the "capacity box" problem. Every AI model has two hard limits: how much text it can hold in memory, and how many logical steps it can chain together reliably. Current interfaces give you zero signal when you're hitting either wall. You just get gradually worse code with no explanation.
What Actually Works (But Tools Won't Tell You)
The practical workarounds Reis describes are telling, because they all require fighting against the tool's own interface:
- Roll back instead of following up. When code is wrong, starting a fresh prompt with corrections produces better results than a back-and-forth conversation. But every tool is designed around conversation.
- Use sub-agents for isolated tasks. Some tools now support delegating chunks of work to separate agent instances that return only a summary. This keeps your main context clean. Most developers don't know the feature exists.
- Write specs before prompting. Maintaining a
docs/folder with architecture decisions and feature specs gives the model structured context. This is old-school software engineering, but AI tools don't nudge you toward it. - Be selective about what tools you expose. Every MCP server or plugin you connect adds to the model's context load. Less is often more.
The uncomfortable implication: the developers getting the best results from AI coding tools are the ones who already know how to write clear specs and break problems into small pieces. The UX isn't closing that gap.
The Real Competition Is in the Interface Layer
Reis's argument lands at an interesting time. Cursor, Windsurf, and Claude Code are all competing primarily on model quality and speed. But if the interface is what determines whether a developer actually gets good output from those models, the real differentiator might be the tool that figures out structured workflows first - guiding users toward spec-driven development instead of freeform chat.
Right now, none of them are doing that well. The tools treat every developer like an expert prompt engineer, which most are not. Until that changes, the gap between what AI coding tools can do and what most people actually get out of them will keep growing.