Every stop was a question

I write ad playbooks for my own products: one document with twenty or so short pieces in it. Five hooks, ten content angles, five calls to action. The ad tool in my Mac app could already make each one into a video, but each video had five to seven approval stops along the way. Twenty pieces meant well over a hundred clicks.

I asked for no stops at all. Drop in the document, pick a style, come back to finished videos.

What a stop actually is

It is easy to think of an approval stop as friction. But each one was the pipeline asking me something: is this the right script, the right length, the right voice, the finished video. Removing the stop does not remove the question. It means the answer has to be decided ahead of time.

So we wrote the answers down. Hooks run three to five seconds, content fifteen to thirty, calls to action five to eight. Hooks and content carry no call to action. Each piece is made on its own and named; when Claude proposed joining pieces into complete ads, I said no, because that is my job afterwards. A video counts as approved when it finishes. Missing tools are checked before the run starts, not discovered halfway through the night.

Written as a table, the automation became something I could review.

Testing the expensive assumption

Before building further, Claude ran the split on my real playbook, with the real model. It came back with exactly twenty pieces. That one test cost a few cents and confirmed the core idea before hours went into it.

It also had its numbers checked. A reviewer reading the design against the code found Claude had left a costly step out of its estimate. Claude came back and told me, rather than keeping the lower figure. I asked whether it was running on my subscription, which reframed the cost as usage against my plan, and we went ahead.

What “safe” means overnight

The bugs the reviews caught were not crashes. A run could get stuck in “stopping” forever. A step could keep asking the same question and spend money each time. A run would have rebuilt a video I had edited by hand, overwriting my change.

When nobody is watching, the worst failures are the quiet ones: things that spend, stall or overwrite without telling you. Designing unattended work means hunting for those first.