Better prompts get credit that architecture deserves. The developers getting strong results from AI assistants are not wordsmithing instructions; they are building codebases with clear boundaries and fast feedback loops, the same discipline that has always separated maintainable systems from fragile ones.
The Details:
Structure teaches by itself. When a function name predicts its behavior, a test confirms it, and a comment explains the reasoning, a model can infer the rules without anyone narrating each decision. If a capable stranger, human or AI, cannot orient quickly in your repo, the system runs on tribal knowledge that no model can access either.
Constraints encode institutional memory. A module that exposes three functions and hides the rest, a type that rejects nulls, a linter rule that blocks the wrong layer: each is a compressed lesson from a past failure. Anyone who bumps into that constraint gets the lesson without repeating the incident, so treating constraints as load-bearing infrastructure rather than friction produces predictable model behavior.
Boring beats clever in configuration. Flat, consistently structured YAML or config files let a model read intent directly. Nested, key-reused, brevity-optimized configs force the model to spend its reasoning budget on archaeology instead of the actual problem, since every fresh context window starts from zero.
Mistakes are diagnostic, not incidental. When a model gets something wrong, a human would likely have struggled at the same spot, just more quietly. The fix is not a better prompt; it is asking what signal was missing from the system and writing it down where the repo can carry it forward.
Bottom Line: Fix the system that let the mistake happen, and the prompt problem mostly disappears on its own.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
