Deliberately Using Less Capable AI Models to Improve Thinking

The author discusses how using less capable AI models can actually improve the thinking process and lead to better design decisions, rather than relying on overly powerful models that can produce working code without much thought.

Details

The author has started deliberately using less capable AI models for certain tasks, not because they are cheaper or faster, but because the constraint forces them to think more deeply about the problem at hand. Overly powerful models like Opus 4.5 can take a vague prompt and generate working code, but the author found themselves defending decisions they hadn't consciously made. In contrast, models like Sonnet require more precise prompts, which forces the user to articulate the problem clearly. This articulation becomes the architecture work. The author has adopted a workflow where they use the most powerful models for exploration and review, but dial it back to constrained models for the initial implementation and refactoring. The key insight is that the prompt should be like a design document, explaining the reasoning behind the task, not just the task itself. The struggle to articulate the problem is the real work, and the friction of less capable models can be a feature, not a bug.