Every Cut

The bottleneck is what we can dream of

I keep saying this in our Friday calls: Whatever we can dream can now be built, so we should dream bigger. I had tried it once already, with a professional audio workstation that I now use every day. A video editor was the biggest thing I could think of next.

I record product demos in Screen Studio, and I keep running out of room in it. Putting several clips together, laying a title on top, mixing music under a voice: each of those sends me to a second editor after an export. And no agent can help in either one, because neither has tools an agent can call. What I pitched to the team was Descript, Screen Studio, After Effects, and Premiere in one.

A video editor is a hard test. Either the frame that plays is the frame that exports, or it is not.

The clock

It did not start from nothing. It started from the template Frontier and our other experiments use, which already had sign-in, agent tools, deploys, and tests. Everything about video was new.

Day 1, October 7. The first commit is at 8:11 in the morning, and the app is called Cutroom. Before 10 it can open a Screen Studio project, record the screen and a webcam, and let an agent do what a person can through one tool path. By 4 in the afternoon it exports an MP4 that matches the preview and cuts a recording by its words. By midnight it has a timeline you edit by hand. Forty-eight changes land on main that day.

Days 2 and 3. I was iterating like crazy and adding a lot: cameras and switching between them, a mixer, a floating recording control, changing a spoken word in the speaker's own voice. By the end of day 3 it is live at cut.every.to for the team.

A Tuesday night. The video editor on our team joined the channel and asked for LUTs, the color tables footage is graded through, grouped by camera. They were in the build the next morning. He then sent a folder of the nastiest footage he had, which we call the torture pack. It found three places where an export did not show the frame that played. So it failed the test, three times. All three are fixed.

The first live test. We opened the studio with three people. The picture worked and nobody could hear anybody. It was fixed the same day, and that afternoon we recorded the first meeting that worked, every camera on its own track.

A week in. A team meeting, held in it.

Day 10. 484 changes on main.

What it cost, and what it is not

Three days to live, a week to get this far. At list prices the tokens for that week come to somewhere between $11,000 and $14,000, about what one engineer costs in a month.

Most features were built and tested by agents in a cloud sandbox, against stand-ins. Every release note says which parts have never touched a real machine, and the docs say it too. So when something breaks I know where to look first.

The strongest pushback came from our video editor, about a feature that copies a reference video down to the pixel. He would rather have the tool name what is going on in a reference and ask which parts you want. He wants a partner and a teacher. I agree. I do not want a prompt that replaces the person. I want professional tools with a person in them, because people are still better at detail and polish.