app

AI Flow

Public Made by Adom

by adom

Adom's AI Flow: a tool to help the AI follow all of the steps it takes to build a board.

kicadpcbedaroutingplacementcopper poursablationcurrent densitythermalkelvincomparison
Overview Files Skills Discover Packages Releases Dependencies Issues10 Breadcrumbs Pull Requests Activity
Closed PR #68 released in 0.2.53 on 5 Nov 2026 PR #88 released in 0.2.69 on 11 Nov 2026 bug report

compose: final video is soft (built from 720p 10x cuts at about 265 kbit/s) and the Fields viewer narration is not captured

Glenn Jones · ·closed by Colby Knox

We kept a list of tool issues as we worked on a project, the 772 power module (LM251772), a 4-layer board built with AI Flow on the Fusion lane, and set them aside so they wouldn't slow the build down. Now that the project is finished, we're filing them together. Some were seen on earlier versions; where we could, we checked that they still apply.

Versions: adom-aiflow 0.2.44, Fusion lane.

What happens (772 power module, LM251772 run):

  • The raw window recordings are 3840x2160 at about 42 Mbit/s. The 10x cuts compose uses are 1280x720, and the final 1920x1080 video came out at about 265 kbit/s video bitrate, so on-screen text (Fusion, the Fields viewer) is visibly soft.
  • The Fields viewer speaks an explanation for each view, but the window recorder records video only, so none of it is in the final video. Compose adds its own TTS, which doesn't describe what the viewer was showing.

Expected:

  • Compose from the raw recordings (or cuts at the output resolution) with a bitrate that keeps UI text sharp.
  • Capture the app's audio for the Fields tour (a desktop recording with system audio, or splicing the viewer's cached narration clips into the timeline).

No design files attached; happy to describe more on request.

2 Replies

Colby Knox marked this in progress ·
Colby Knox mentioned this in pull request #68 ·
Colby Knox ·

Glenn, thanks. The soft video is fixed in adom-aiflow 0.2.53.

Why it was soft: the clip cuts were 1280 wide, and compose squeezed any video over 10 MB to fit. That 10 MB wasn't a wiki limit, it was AI Flow's own, and it's where your 265 kbit/s came from.

What changed:

  • The cuts are now up to 1920 wide, and the final 1080p video is encoded for quality.
  • Only a file over 90 MB is re-encoded to fit, because one upload to the wiki tops out at 100 MB.
  • The overlays keep their size and place.

What to expect: bigger files. A real 117 s video went from 7.6 MB to 37.5 MB, with small on-screen text visibly sharper, and the 10x cuts about double. Your own run may differ, since a long or busy video takes more.

Older runs: run recut, then compose, to get the sharper cuts. recut needs the raw recordings in the run folder, so bring back any you moved. Older cuts still compose, just at their old sharpness.

Not tested: pushing a video this size to a wiki page. The largest I've seen on a page is about 13 MB, so please tell me if yours is refused.

Still open: the Fields viewer's spoken explanations aren't in the video, because the window recorder captures picture only. I'm leaving this issue open for that, but it isn't being worked on right now.

Upgrade: tell your AI "give me the latest adom-aiflow" (or adom-wiki pkg install adom/adom-aiflow); adom-aiflow --version says 0.2.53.

Colby Knox removed the in progress marker ·
Colby Knox marked this in progress ·
Colby Knox linked pull request #88, which closes this when it merges ·
Colby Knox closed this in pull request #88 ·
Colby Knox ·

Glenn, the last part of this is in AI Flow 0.2.69: the Fields viewer's explanations now reach the video.

Adom Bridge's window recorder takes picture only, so the viewer's voice can't be recorded with the clip. The video uses the viewer's own words instead:

  • tour fields keeps what the viewer explained during the take on the clip. That's each view's text and its mp3, from the Fields folder of the run, in the tour's order.
  • compose says those words over the Fields segment, in the same voice. They carry this board's numbers: the busiest net and its density, the hottest part and its rise. It uses the fullest version that fits the segment (two views' explanations, then the first sentence of three, then one view's), and only falls back to its generic line when the viewer said nothing.

Runs made before this read the folder directly, so a compose of an older run picks them up too.

Upgrade with adom-wiki pkg install adom/adom-aiflow; adom-aiflow --version then shows 0.2.69.

Log in to reply.