What part of AI-assisted software development still feels surprisingly manual?
I've been building with tools like Claude Code, Cursor, and Replit, and they've dramatically sped up coding.
But I've noticed there are still parts of the workflow that feel surprisingly manual.
For example:
Finding a prompt that worked a few weeks ago.
Remembering how a specific feature was generated.
Keeping architecture and AI context organized across different tools.
Iterating on AI-generated code without starting over.
I'm curious what everyone else's experience has been.
If you're actively building AI apps or AI-powered products:
What's the most frustrating part of your AI development workflow today?
How are you solving it?
Is there a workflow you've simply accepted because there's no better solution?
I'm asking because I'm trying to better understand how builders actually work, I'd love to learn from your experiences.
Replies
The most manual part for me is preserving context between sessions. The code survives, but the reasoning usually doesn’t: why a decision was made, which constraint mattered, what the agent already tried, and what should not be changed.
A simple fix is keeping a small repo-level memory file with architecture decisions, active constraints, failed approaches, and a short checkpoint after each task. It works even better when the agent updates it automatically before finishing. That makes switching between Cursor, Claude Code, or another tool much less like starting from zero every time.
"remembering how a specific feature was generated" is the one that gets me. I'll re-explain the same constraint to an agent three times in a week because it lived in my head instead of somewhere it could be pointed to.
For me it's iterating without starting over - ask for a small fix and the agent rewrites half the file or touches things that were already working fine. I've started forcing much smaller diffs and reviewing every change before accepting, but it still feels like babysitting. Also relate hard to what Alper said - the reasoning behind a decision disappears way faster than the code does.
@harafernando
Keeping architecture and context organized across tools is the one that never fully goes away for me. The code survives between sessions, the reasoning behind it usually doesn't. Writing decisions and constraints somewhere the agent can be pointed back to helps more than any prompting trick, it turns personal memory into something the whole project can rely on. Curious if anyone's found a lightweight way to keep that file updated without it becoming a chore itself.