autopilot takes a typed idea all the way to a finished video

the whole promise of spoolcast, working unattended. you type an idea in one line and walk away, and it comes back with a rendered video.

the design decision that matters: autopilot has no private path through the product. it calls the same engine actions the interface calls, presses the same buttons a person presses, step after step. that is why a run cannot quietly diverge from what you would get by clicking. it fills in what you left empty so a bare idea runs through, counts the steps the project actually has instead of assuming a fixed workflow, and says what a run will cost before it starts.

getting there took most of two weeks of small failures. autopilot not getting past its first step. the core message step saving but not continuing. the start button appearing on steps it could not actually work from. none of it interesting on its own. all of it required before the thing could run alone.