Getting Claude/Codex to build to plan?
Maybe it's just me and my setup or approach, but I'm finding repeatedly that I spend a couple of passes on a planning document, lay out technical specs, define success/failure criteria, test requirements, outline automation testing requirements, etc. and yet...the outputs are husks of what I designed. Loads of stuff isn't done at all, a bunch is just stub code, no tests are written, no acceptance criteria was ever done, etc. I think I've internalized the approaches Jake evangelizes, and I've got it all in documents, but at the end of the day I'm finding (via Claude's own 'admissions') that the models just treat any rules and guidance as optional. They just ignore the requirements to read the documents - e.g. it will say stuff like 'The instructions said to read the specs, I read the first few lines and then figured I knew what to do so I stopped and got to coding' or some variant of that. I ended up building a bunch of harness gates and hooks (in Claude) but now I find that Claude just spends more time figuring out how to satisfy the hook validation than doing the actual work needed to pass it. I can build simple sites or tweak codebases with super small targeted changes, or build remotion videos to my hearts content using the processes Jake is showing, but anything substantive? Nope. It fails miserably. I want to believe I'm just doing it wrong, but so far I'm not seeing how it would be at all possible to do any kind of substantial coding using these coding tools. I'm a CLI person, always hand coded (the hard way, I know) so maybe there are better answers using IDEs? Would love any thoughts others have on how to avoid these issues and get the models to actually follow the file structure requirements and not just glance at the docs but read them and follow them.