I built five little robots to stop me lying

Working with Claude all day, the failure isn’t that it’s dumb. It’s that it’s confident, and so am I, and confidence at the finish line walks you straight past the moment you were supposed to check. I declared one refactor done three separate times. It was wrong three separate times. So I started building gates.

There are five, and each one exists because of a specific time I got burned. One won’t let me claim a ticket’s status without a receipt from the actual source, because I once put three false facts straight into a stakeholder doc. One blocks “ready for QA” unless it came from the real readiness script, because I once said five tickets were ready when the script said zero. One blocks a Jira comment that skipped the voice pass. One stops two terminals running the same job at once. One won’t let me type “done” without proof something ran.

The gates are deliberately stupid. They don’t read my work and judge it. They check for a little receipt token, that’s it. The point is forcing a conscious decision at the exact spot I’d otherwise sail past. They also fail open, so a broken gate lets the work through instead of blocking everything. A safety rail that turns into a roadblock gets ripped out by lunch.

My favorite watches the others. When I save the same lesson to memory for the fourth time, it tells me to stop writing notes and build a gate instead.

I automated my own conscience because the manual one kept calling in sick.