gbro-collage-broll: Turn a 5-Second Voiceover Into a Premium Paper-Collage B-Roll Clip — With a 3-Gate Cost-Control Workflow
Short-form creators know the pain: you’ve got a punchy voiceover line, but the visuals under it are generic stock footage or a slow Ken Burns zoom. gbro-collage-broll attacks exactly that gap — and does it with an unusually disciplined, cost-aware design.
The pitch: feed it a ~5-second voiceover line, and it compresses that line into a single sharp visual metaphor, then generates a premium editorial halftone paper-collage B-roll clip — flat solid-color paper backgrounds, black-and-white halftone photo cutouts, and colored cardstock accents that slide in and assemble stop-motion style (not a lazy fade or zoom). Output is a 9:16, 5-second, 720×1280, 24fps silent MP4, ready to drop straight under your narration. It’s powered by Gemini Omni Flash first/last-frame video generation.
The real innovation isn’t the prompt — it’s the gates
Here’s what makes this project genuinely smart. The author states plainly that the core of the skill isn’t a prompt template — it’s a mandatory three-gate approval workflow designed to keep you spending your attention on aesthetic judgment rather than burning generation fees:
- Gate 1 · Metaphor confirmation. It outputs only the visual-metaphor plan — core meaning, key objects, background color, assembly order. No images, no video. Editing text here is free.
- Gate 2 · Still-frame confirmation. Only after you approve does it generate the colored collage still plus a contact sheet, then stops and waits again. Regenerating one image is far cheaper than re-rendering a clip.
- Gate 3 · Video generation. Once the still passes, it runs
gemini-omni-flash-previewfor the first/last-frame assemble animation — complete with QA: per-second frame sampling, empty-background first-frame verification, and last-frame comparison.
In batch mode it even supports partial approval — only the lines you’ve confirmed advance to the next stage. This is the kind of guardrail most “AI video generator” tools skip entirely, and it’s the difference between a fun demo and something you’d actually run on a client budget.
The aesthetic
The look is deliberately editorial: strongly flat pure-color paper color-fields, high-contrast halftone photo clippings, and pops of colored paper — everything animating in with a tactile, physical stop-motion feel. The demo GIFs show themes like “collaborators pressing sci-fi film out of a machine,” “a printing press mass-amplifying an error,” and “a director’s hand arranging chess-board positioning.” It’s a strong, coherent visual identity rather than generic AI slop.
Setup and use
It runs as an agent skill (Codex or Claude). Install by cloning into your skills directory:
git clone https://github.com/pyang5166/gbro-collage-broll.git ~/.agents/skills/gbro-collage-brollYou’ll need a Codex environment (Gate 2 uses the built-in image_gen tool), a GEMINI_API_KEY from Google AI Studio (video is pay-as-you-go), Python 3.10+, google-genai >= 2.10.0, and ffmpeg/ffprobe. On first trigger the skill runs check_setup.sh and tells you what’s missing. Then just say:
collage b-roll: <your ~5-second voiceover line>…and step through Gate 1 → 2 → 3.
The takeaway
There’s a broader lesson here worth stealing regardless of whether you make collage B-roll. Most generative-media tools optimize for one-click magic, which quietly optimizes for wasted spend — you generate a full video, hate it, and pay for it anyway. gbro-collage-broll inverts that: it makes the cheap decisions (words, then a single still) the checkpoints, and only spends real money once you’ve committed. That “confirm before you burn” pattern is good design for any pay-per-generation AI pipeline.
If you produce vertical short-form and want a distinctive, premium visual style with predictable costs, this is a sharp little tool — and a nice template for how to wrap expensive generative models in human-in-the-loop discipline.