one shotting/"context engineering" larger apps or bigger code changes with claude code gives me fragile results. common issues:

- too much code gets generated. flows might work, but the code is slop.
- many UI/UX bugs. every iterative fix dirties the code further.
- claude invents new features/abstractions that I never asked for

what's working better for me is acting as the architect. I break down the task myself into what would usually be 15-30 minute sessions if I were to write the code manually.

- I ask claude to implement each increment
- I test it and give feedback to claude
- I then review the code and ask for any necessary refactors
- git commit -m "checkpoint"
- rinse and repeat

this places me in the role of PM/tech lead on a team. results in better UX and more resilient code.

exploring. prev: @clankfun, improbable, github. https://ntik.me

Found a similar approach to be quite helpful when approaching coding stuff with AI. 

I really don’t like when the AI goes off and starts creating new files that I haven’t named or planned beforehand. Get super messy. 

Been doing front end stuff, and found actually creating page routes, creating components and functions but leaving them blank at the beginning has been really useful. 

It tends to respect my “blueprint” - this upfront work saves me massive headaches trying to understand what it’s done - but can easily ask if there is a better way to structure or approach a feature.