Non-developers hit a wall at 65 percent of an AI-built app: what actually helps finish it?
A non-developer building with Claude Code and Codex describes stalling at roughly two-thirds done, with the model ignoring agreed requirements and changing things unasked despite rules files and CLAUDE.md. The likely cause is the project outgrowing what fits in the model's working memory, so it improvises. The practical answer is smaller jobs: one feature at a time with a short written spec, tested after every step.