AI Flow: Video

@@ -4,7 +4,7 @@ version = 4  [[package]] name = "adom-aiflow-video"-version = "0.7.0"+version = "0.7.1" dependencies = [  "clap",  "serde",

@@ -2,7 +2,7 @@ # Copied from adom/adom-aiflow template/flow-app. This copy is yours: change anything, never depend on aiflow. [package] name = "adom-aiflow-video"-version = "0.7.0"+version = "0.7.1" edition = "2021" license = "MIT" 

@@ -104,4 +104,4 @@ Your film is yours: it carries no Adom logo. The Adom logo sting is only for off  ## Status -0.7.0. It grew out of the Ghost Hand films ([john/ghost-hand](https://wiki.adom.inc/john/ghost-hand)). Working today: the flow and its hints, clip recording, the disk estimate, script feedback, the text privacy scan, file tracking and cleanup, the walkthrough director, the narration page (voice, camera picture, drawing, playback check, AI notes), and the YouTube channel check, title and thumbnail feedback and upload. Still done by the AI by hand, with the skill as the guide: the cut itself, putting the voice, picture and drawings on the film, lining up a second recorder, composing thumbnails, and the frame-by-frame privacy check.+0.7.1. It grew out of the Ghost Hand films ([john/ghost-hand](https://wiki.adom.inc/john/ghost-hand)). Working today: the flow and its hints, clip recording, the disk estimate, script feedback, the text privacy scan, file tracking and cleanup, the walkthrough director, the narration page (voice, camera picture, drawing, playback check, AI notes), and the YouTube channel check, title and thumbnail feedback and upload. Still done by the AI by hand, with the skill as the guide: the cut itself, putting the voice, picture and drawings on the film, lining up a second recorder, composing thumbnails, and the frame-by-frame privacy check.

⋯ 1 unchanged line ⋯

⋯ 1 unchanged line ⋯

⋯ 1 unchanged line ⋯

@@ -1,7 +1,7 @@ {   "slug": "adom-aiflow-video",   "type": "app",-  "version": "0.7.0",+  "version": "0.7.1",   "title": "AI Flow: Video",   "description": "Make a narrated walkthrough video of a board built in Adom Hydrogen, put it on the project's wiki page and help the human upload it to YouTube. A dynamic skill: a small Rust binary that suggests each step, records clips, ships a ready-made 3D walkthrough director and gives gentle feedback; the AI does the work. Open source (MIT): the full Rust source ships on this page; build it yourself with ./build.sh.",   "category": "Developer Tools",

@@ -1,7 +1,7 @@ {   "slug": "adom-aiflow-video",   "type": "app",-  "version": "0.7.0",+  "version": "0.7.1",   "title": "AI Flow: Video",   "description": "Make a narrated walkthrough video of a board built in Adom Hydrogen, put it on the project's wiki page and help the human upload it to YouTube. A dynamic skill: a small Rust binary that suggests each step, records clips, ships a ready-made 3D walkthrough director and gives gentle feedback; the AI does the work. Open source (MIT): the full Rust source ships on this page; build it yourself with ./build.sh.",   "category": "Developer Tools",

@@ -137,3 +137,5 @@ The thumbnail's few words are a promise, and the narration has to keep it, so se A person pointing at the product beats a face in a circle. `narrate pose "<what to do>" --thumbs <product picture>` opens the camera on the narration page with the product's own picture on the view, framed in yellow under a large "POINT AT THIS" marker (they can drag it; without --thumbs it is a plain dot, which people find hard to read, so always give the picture), counts down and takes three stills, saved as `pose-N.jpg` in the takes folder, mirrored the way they saw themselves so what they pointed at is where they point. Then `thumbkit --out thumbkit/` writes two helpers: `cutout.py` cuts them out of their background (a person-segmentation model run with onnxruntime, offline) and `compose.py` places them pointing at the product beside a two-line headline. Look at the result, adjust, and show two or three in the picker. Pick the still where they look at what they point to. If the page says another app is using the camera, tell them: the browser cannot take it from another program.  Only one open copy of the narration page acts on camera, microphone and recording commands: a hidden tab never does, and among visible ones it is the newest or the one the human last touched. `narrate status` shows `pageId`, `inControl` and `hidden` for the page that last reported.++Thumbnail checks people have asked for, every time: the finger lands on the product itself, not on something beside it; whatever makes the product make sense stays in the picture (a faucet keeper needs the faucet in view); the headline stays large (about 110 to 120 px at 1280 x 720, two or three short lines); and there is clear space, about 80 px, between the headline and the picture. If one of these forces another out, take a new picture of the product from a better angle instead of cropping.