How I Turned One Open‑Source B‑roll Skill into Five Distinct Video Styles

First look: same skill, five B‑roll styles

I tested two vertical styles on the same voice‑over sentence.

想日更，光剪一条视频就能把整个晚上耗光。

The first version is the upstream skill’s strength: 编辑半调拼贴 . Characters, film strips, sun and moon slide in from solid‑color paper, suitable for tech and opinion narration.

The second version swaps the metaphor for a copier‑style Zine, using dirty white paper, black toner, tape and staples for a rougher mood.

Next I used the same passage from "Journey to the West" to test three horizontal styles.

火焰山挡住前路，一把芭蕉扇，让蛮力第一次向智慧让路。

剪影镂空纸艺 uses layered paper cut‑outs to depict the mountain, flames, fan and pilgrimage team, fitting myths and story explanations.

档案卷宗 turns the sentence into an evidence chain of old photos, route maps, red lines and pins, suitable for case studies and historical material.

超现实摄影蒙太奇 stitches impossible scale relationships together, ideal for abstract viewpoints, counter‑intuitive statements and strong emotions.

Changing the material and motion grammar of the same sentence yields completely different visual impressions, which is why I expanded the skill: B‑roll should not be limited to a single paper‑collage template.

Skill core: turning script into visual metaphor

The upstream project is gbro-collage-broll. After receiving a script, it forces three gates instead of calling a video model directly.

Gate 1: confirm the visual metaphor and style for each sentence.

Gate 2: generate a static frame and confirm characters, objects, composition and material.

Gate 3: after the static frame passes, generate a 5‑second video from the start and end frames.

I kept this workflow and added eight style presets.

编辑半调拼贴 – suitable for technology, opinion (tested)

几何剪纸 – suitable for product, abstract relations (to be tested)

Riso 套色印刷 – suitable for culture, creative expression (to be tested)

复印机 Zine – suitable for satire, repetitive labor (tested)

档案卷宗 – suitable for history, case, character relationships (tested)

超现实摄影蒙太奇 – suitable for counter‑intuitive, emotional viewpoints (tested)

蓝图结构图 – suitable for principles, systems, tutorials (to be tested)

剪影镂空纸艺 – suitable for myth, story, cultural content (tested)

These three screenshots are Gate 2 static frames for the first three sentences; changing metaphor, style or composition is still possible.

Step 1: Install the open‑source skill into an Agent

Repository: https://github.com/pyang5166/gbro-collage-broll

In an Agent that supports Skills, enter the prompt:

Prompt: Install this skill and create the related video: https://github.com/pyang5166/gbro-collage-broll

The agent automatically clones the project and reads SKILL.md. Manual installation example:

git clone https://github.com/pyang5166/gbro-collage-broll.git \
  ~/.agents/skills/gbro-collage-broll

The original project required a GEMINI_API_KEY, but I avoided adding a new API by asking the agent to reuse existing image/video capabilities.

Agent reuses the current image generation for Gate 2, then connects Gate 3 to a video model. FFmpeg normalises resolution, frame rate and strips the audio track.

Before generation, the agent prepares a clean first frame and assembles it with the confirmed last frame. Below are the three actual blank first frames.

Step 2: Feed the full script and let the agent split it

I gave the agent the complete voice‑over script without rewriting it. The agent decides which sentences deserve B‑roll and which to skip, outputting a table with original text, a generate flag, visual metaphor and suggested insertion point.

Pure transition sentences, already‑covered live actions, or overly long sentences are skipped. Only sentences that can be expressed as a clear visual premise are kept.

For this demo the first three sentences were split into three visual concepts: a tiny script paper crushed by a huge editing tool; the creator trapped in a loop of searching for material, matching visuals and adding effects; the sun swallowed by a film roll and emerging as a moon.

After confirming the split, I sent a single prompt:

Prompt: Generate all frames using the “复印机 Zine” style.

Agent produced three static frames; I asked it to redo the first one because the scissors were too close to the edge. After all static frames passed, I replied “through” and the video was rendered.

Step 3: Review the final clips and insert them into the main video

I check four aspects: whether the video assembles from an empty canvas piece by piece, whether extra characters or fake text appear, whether the style drifts, and whether the final frame matches the confirmed static frame.

Comparison of the Gate 2 static frame (left) and the video end frame (right) is far more informative than a vague “looks good”.

During testing I encountered a 429 rate‑limit; tasks without an ID were retried after concurrency dropped. Do not re‑run batches that have already passed. Minor paper curl or object contraction that does not affect semantics does not need to be redone.

The generated MP4s have no audio, which works well for voice‑over B‑roll. After recording the live video, drag the B‑roll clips onto the timeline, align them with the corresponding sentences, keep the original voice‑over, and add subtitles, background music and SFX as needed. The agent does not produce the full final video but removes the most time‑consuming parts of material search, visual planning and basic motion effects.

In summary, the full script is first given to the agent for splitting; only visual‑izable sentences proceed; the workflow is visual metaphor → static frame → video, yielding ready‑to‑insert B‑roll assets.