Blog

This article looks at a session where the AI started by proposing the logging design before touching code. That approach worked well, but two misses still made it through: a redaction rule that didn’t cover the main case

A sidebar property switcher and a background job runner both worked at the feature level. The trouble showed up in everything around them: branch workflow, Playwright selectors, and test environment setup.

After several successful feature layers, one session lost track of the actual codebase state and started patching against memory instead of files. This post explains the warning signs and why starting fresh was the right

Claude got me from zero to a working SaaS skeleton in one session, but it completely missed the part that

A practical look at bananas and coding, with clear context and real implementation details.