Step 1 · The input

Everything the engine was given

One form, two screenshots, one command. This is the entire human contribution to the run.

The intake form, as I submitted it

The engine's front door is a local form that writes a run folder. Below is this run's form with exactly what I typed, typos included ("Cluade" is mine). The right-hand pane shows the idea.md it generates live.

The intake form filled in: run name, the two-paragraph idea, target product Claude, two screenshot thumbnails, and the generated idea.md preview — The intake form for this run. Only two fields are required: the idea and the product. No URL was given, so the screenshots carry all the visual evidence.

The two screenshots I uploaded

the only visual evidence the agents got

Claude Android app with the navigation drawer open: New chat, Chats, Projects, Artifacts, Code, Dispatch, then Starred and Recents — `Screenshot_20260713-180822.png`
The drawer. Code sits 5th, Dispatch 6th.

Claude Android app home screen: greeting, chat composer, model chip — `Screenshot_20260713-181535.png`
The chat home the app always opens to.

What landed on disk

"Save run folder" wrote this input folder. That is the entire specification the pipeline starts from.

inputs/cluade-mobile-app-menu-reorder/

├─ idea.mdthe two paragraphs above, plus "why now" and constraints24 lines

└─ screenshots/the two PNGs above2 files

Then one command in Claude Code, from the engine's repo:

$ /propose inputs/cluade-mobile-app-menu-reorder

From here the pipeline runs its five stages in a fixed order: context, discovery, definition, risk sweep, synthesis. Every artifact in the rest of this walkthrough was written by an agent during that run and is shown unedited.