AI Is Very Good at Implementing Bad Plans

The developer says AI coding tools often produce polished code from flawed plans, so the real risk is bad architecture and assumptions, not syntax.

To catch those errors earlier, they now send implementation plans to multiple independent models, compare the critiques, and revise the plan before writing code.

They also added a practical rule: when tool output gets too large, write it to files instead of pushing against response limits.

The takeaway is a low-friction workflow for plan red-teaming that can prevent costly mistakes in AI-assisted development.