Assembling Aesthetics: A Three-Gate Skill for Halftone B-Roll

The Anti-Template Aesthetic

Most AI video tools chase a cinematic look. Platforms like Gling and OpusClip promise instant B-roll—forests for nature documentaries, smiling kids for charity pitches, generic stock motion for any script under the sun [8][11]. The results are smooth, plausible, and often forgettable. The repository gbro-collage-broll takes the opposite tack. It commits to a single, highly specific visual language and refuses to be general purpose.

The aesthetic is editorial halftone paper collage. Flat, saturated color fields—deep purple, mustard yellow, crimson, teal—serve as backgrounds. Black-and-white halftone photographs, clipped and pasted, form the informational layer. Accents of colored card stock provide punctuation. The motion is not a fluid camera push or a slow cross-dissolve; it is a staccato sequence of elements sliding into frame and locking into position with the slight imprecision of stop-motion. The README calls this an “assemble-from-empty” animation, and the default deliverable makes the use case explicit: a five-second, 9:16, 720×1280, silent MP4 at 24fps. This is not a standalone film. It is a visual exclamation point designed to sit beneath a short-form voiceover.

The halftone look itself is not novel. It has persisted in stock media libraries [3][9], video templates [6], and mobile filter apps [12] as a shorthand for print-era authenticity and pop-art punch. What is new is the automation of its assembly. The skill does not merely apply a halftone filter to existing footage; it generates the entire collage—objects, background, layout—from a single sentence of voiceover text, then animates the construction. In a landscape where AI video defaults to cinematic parallax because that is what dominates training data, this is a deliberate act of stylistic narrowcasting.

First Frame Empty, Last Frame Full

The technical mechanism relies on a specific affordance of Gemini Omni Flash: first/last-frame video generation. Rather than prompting the model to imagine a “stop-motion assembly” and hoping the temporal coherence holds, the skill externalizes the problem. It generates two static images—an empty field for the first frame, and a fully composed collage for the last frame—and asks the model to interpolate the motion between them. The result is a clip where paper elements appear to fly in, settle, and snap into place.

This approach is pragmatic rather than magical. Stop-motion animation is notoriously difficult to prompt in pure text-to-video systems because it requires discrete, non-physical motion. Objects must slide across surfaces without friction, blur, or obeying real-world gravity. Diffusion models trained on natural footage tend to warp, morph, or apply unwanted physical dynamics when asked to make paper move like paper. By constraining the task to a start-state and an end-state, the skill turns an open-ended generation problem into a frame-interpolation task. The model handles the in-betweening while the human retains control over the composition.

The README is explicit about rejecting default video aesthetics. The elements must not fade in or execute a slow zoom. They must slide, snap, and lock. That rigidity is a feature. It protects the output from the lazy cinematic language that most video diffusion models default to when given an ambiguous prompt. The five-second duration matches the average breath of a short-form sentence; the silence acknowledges that the clip is a substrate, not a soundtrack.

Three Gates Before You Burn Cash

If the aesthetic is the skin, the three-gate workflow is the skeleton. The README states plainly that the core of the skill is not prompt templates but “mandatory three-stage approval.” Gate One is metaphor confirmation—text only, zero API cost. The agent proposes a visual metaphor, lists the key objects, assigns a background color, and defines the assembly sequence. The user approves or revises. Gate Two generates a static collage frame and a contact sheet; this uses the host agent’s built-in image generation, which is cheaper and faster than video. Only after the static image passes does Gate Three trigger the Gemini Omni Flash video render, complete with a QA checklist: per-second frame extraction, first-frame empty-field verification, and last-frame comparison.

This is a direct rebuke to the “generate and pray” model that dominates consumer AI video tools [8][11]. Those platforms optimize for velocity: upload a script, receive B-roll seconds later. The trade-off is genericness and waste. A bad metaphor rendered in 4K is still a bad metaphor, and video generation APIs are priced per second or per job. Burning credits on a clip with the wrong concept is economically painful for solo creators. By inserting hard human checkpoints before the expensive video stage, the skill shifts the cost curve from generation to curation. The README frames this in blunt economic terms: changing text at Gate One is free; regenerating an image at Gate Two is far cheaper than rerunning a video at Gate Three. In batch mode, the user can approve individual lines selectively, ensuring that only the strongest concepts reach the video tier.

The contact sheet in Gate Two serves as a storyboard. The QA in Gate Three treats the output as a deliverable that must be inspected, not a toy. This mirrors traditional animation pipelines, where animatics precede final render, but compresses that discipline into a workflow that runs inside an agent chat.

Agent Skills and the New Craftsmanship

The skill is not a web app. It is a directory of Markdown, YAML, and Python scripts meant to be dropped into an agent’s skills folder. It expects a Codex environment for image generation, a Gemini API key for video, Python 3.10, ffmpeg, and a shared virtual environment. This architecture matters because it signals a shift in how generative media tools are being built. Instead of monolithic SaaS platforms that try to own the entire pipeline, the project is a narrow, interoperable wedge. It delegates image generation to whatever engine the host agent provides, video generation to Google’s API, and quality assurance to standard Unix media tools.

This is a tool for people who already know what a contact sheet is, who understand why a 9:16 aspect ratio matters, and who can inspect a frame extraction for errors. It sits in a growing middle ground between consumer auto-B-roll generators and professional motion-graphics suites. The presence of a SKILL.md protocol file, an agents/openai.yaml interface configuration, and an evals/evals.json file for gate behavior testing suggests the author thinks about this as a production system, not a weekend hack. It reflects a broader trend where creators are packaging aesthetic knowledge as agent-native software rather than standalone applications.

The Limits of Honest Glue

For all its workflow elegance, the project is fundamentally a sophisticated wrapper. It does not train models; it orchestrates them. The FAQ is refreshingly candid about this. A sliver of paper might show at the first frame edge; if you need a rigorously empty start, you will still need a timeline editor. The video model is locked to gemini-omni-flash-preview unless the user explicitly overrides it. The entire value proposition rests on the assumption that Google’s first/last-frame feature remains available and affordable.

These are not fatal flaws. They are the honest boundaries of glue code done well. In the current AI landscape, the competitive advantage is increasingly moving away from model ownership and toward taste architecture—the ability to build systems that enforce creative standards before the meter starts running. The three-gate protocol is a template that could outlive the specific models it currently calls.

Where the field is flooded with tools promising instant cinematic footage, this skill argues for slowness at the right moments. It treats generative video less like a lottery and more like a print shop: lock the layout, approve the plate, then run the press. For short-form creators tired of algorithmic sameness, that discipline may be the most premium feature of all.