Too many clicks: why I wanted a film factory

I wanted to make films. The first obstacle was the work between the idea and the film: entering requests, waiting for clips, finding downloads, checking files and arranging an edit. On 13 August 2026, I asked for a factory that could carry those steps without making me sit beside it.

The first plan concerned Midjourney clips. The route being considered had no general API for the proposed work, so the conversation turned to a browser extension and computer use. I was explicit about pace. I wanted the ordinary clerical work handled, with one job at a time and enough delay to avoid behaving like a flood of requests. The point was to remove repetition from my day.

Codex was to interpret direction, plan shots and judge candidates, while the extension would carry out the clicks and local software would remember what had been requested and what had happened. That division mattered because a browser can close after a click, leaving the next process unsure whether it should try again. A remembered job could prevent a second submission from being mistaken for recovery.

The film belonged in the request

I asked for one hundred questions about the extension, then a phased implementation plan. When Final Cut Pro entered the conversation, I corrected the scope. The factory should assemble the film programmatically. Handing me a collection of clips and asking me to perform the edit would leave part of the original job unfinished.

The plan used FFmpeg, a program for cutting and combining media, with a companion inspector checking the exported files. The archived remit retains the practical boundaries:

- One leased Midjourney tab and one remote job at a time.
- Human-slower interaction pacing and a twenty-minute post-verification cooldown.
- Durable queue, lease, recovery, evidence, and provenance state.
- Automatic media verification, contact sheets, candidate selection, and continuity review.
- Bounded automatic regeneration and re-editing.
- Programmatic picture editing and supplied-audio mixing.

Public source excerpt. Repository snapshot, requiring repository access.

Those lines describe the intended operation. They are a preserved piece of the plan, rather than proof that the complete route ran. They show how early the request already joined generation, verification, editing and sound.

The implementation grew substantial local parts: an extension, a small native host on the computer, a database of projects and runs, a model connection, and a portal for creative direction and references. Tests using deliberately artificial media produced actual master and viewing files. The encoder had done real work on fixture inputs.

Where the route stopped

The external handoff remained harder. A job reader could identify links on a Midjourney page, but its strict matching rules also demanded an exact submission time, complete parameters and an action type. The inspected page did not expose every required field. Codex recorded missing values, and that stage stayed blocked.

Other checks showed an account or runtime was available without completing the actual planning request that the relevant stage required. An authenticated page, a readable job link and a synthetic export each established something useful. Their combination still did not establish the complete direction-to-generated-clip-to-film route. The archived software never received its final acceptance in the inspected history.

That distinction became a recurring concern in the project. I had asked for the work to reach a film. A component could be healthy while the requested sequence remained unfinished. Later short films used a different production route, so their existence could not quietly close the old extension project.

The original brief remains a useful way to read everything that followed. As the factory acquired drawings, scenery, puppets and a whole theatre, its purpose stayed concrete: carry the production work through, and leave me with something I could watch.