My side projects keep following the exact same arc. I spot a real need, get excited, and dive in. Usually I'm building something AI-native, though lately I've been experimenting with physical e-commerce too.
I come in knowing the space and knowing the software patterns I like. Then, a few iterations deep, I have to kill the project because I've built myself into a corner.
Sometimes the AI leads me right into the trap. While building an MCP recently, Codex decided to spin up a local "helper" utility to bridge Codex and Claude. What it didn't mention was that this little helper would now need testing and support across Linux, Windows, and Mac.
After hitting that wall, we wrote a distribution playbook. For this project, the guardrail was to keep the MCP strictly over HTTP. I've attached the playbook below if you're curious.
I used to see these false starts as failures. Now I treat them as probes, an essential scouting phase before the real build.
I run the probe, find the landmines, scrap it, and start v2 with strict guardrails baked right into my intent.md.