How to make AI b-roll for your video

For AI agents: find the right skill in one call — curl -s "https://skills.varg.ai/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://skills.varg.ai/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page.

Cutaway footage for every mention in a voiceover or interview: generated clips, collage or stock-style shots, cut in on time. A step-by-step tutorial for Claude Code, Codex and Grok.

Output: Your video with b-roll inserted, or a folder of clips with timestamps Price: in credits on varg, with a cost estimate before anything runs; the free skills below install in one command.

Recommended route

Best skills for this

  1. varg AI (varg-ai, varg, live) — Start here: Generates each b-roll shot (Seedance 2.5, Kling 4, Veo 3.1, FLUX 3) on one key
  2. Remotion Create (remotion-create, Remotion, live) — Start here: Your own product UI animated as b-roll, from a design-system file
  3. Halftone Collage B-roll (gbro-collage-broll, pyang5166, beta) — Alternative: Editorial paper-collage b-roll from one line of script · Needs: Gemini API key (GEMINI_API_KEY)
  4. HyperFrames Media Use (hyperframes-media-use, HeyGen, live) — Alternative: Finds or makes assets for a HyperFrames edit · Needs: HeyGen CLI and account (optional, paid media)
  5. Watch (video analysis) (claude-video-watch, Brad Bonanno, live) — Adds: Reads the video's transcript and frames to find where b-roll belongs
  6. ElevenLabs Speech to Text (elevenlabs-speech-to-text, ElevenLabs, live) — Adds: Word timestamps, so each insert lands on the word that names it · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)
    • Claude Code: npx skills add elevenlabs/skills --skill speech-to-text -a claude-code
    • Codex: npx skills add elevenlabs/skills --skill speech-to-text -a codex
    • Cursor: npx skills add elevenlabs/skills --skill speech-to-text -a cursor
    • Gemini CLI: npx skills add elevenlabs/skills --skill speech-to-text -a gemini-cli
    • OpenClaw: npx skills add elevenlabs/skills --skill speech-to-text -a openclaw
    • Details: https://skills.varg.ai/skills/elevenlabs-speech-to-text.md
  7. Color Grade AI (color-grade-ai, Isaac Rowntree, beta) — Adds: Matches the grade across real and generated shots
  8. Splice AI clips into a video with FFmpeg (varg-ffmpeg-mixing, varg, beta) — Adds: Splices clips into the original without audio drift

Ask the person first

Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it.

B-roll style?

How much?

How to make AI b-roll for your video: step by step

  1. Transcribe at word level, state the thesis, and write one line per beat: what is it actually about?
  2. Route each beat: named thing to real or your own footage, your product to your UI as motion, number or steps to a data card, idea to collage, a gap to a generated and labelled clip; drop beats with no accurate visual
  3. Write one style spec (palette, type, grain) and pass it to every shot
  4. Stills first on a contact sheet, approve, then animate
  5. Place each insert silent and full-bleed 0.2-0.5 s after the word, check a frame grid of every insert, tag realistic AI shots and export

What works now (researched 2026-10-27)

B-roll that works in 2026 is specific to the exact line being said: your own product screens rebuilt as motion graphics, editorial collage for ideas, real photos for named people and places, and generated clips only for what you could not film, labelled when realistic. Generic stock-style AI clips under an AI voice is the pattern YouTube's inauthentic-content enforcement now demonetizes, and its recommendations favor a real face on camera.

  1. Choose the b-roll type by the line: a metaphor or opinion gets collage or illustration; a number, step or comparison gets a data graphic; a product gets your real UI as motion; a named person or place gets a licensed real photo. Generate realistic footage only for what you can't film. Eian's rule for his two b-roll skills: collage for opinions, feelings and metaphors; riso data graphics when the line has numbers, order or comparison, so the number on screen matches the script. Together they cover about 99% of a talking-head script's b-roll. (Eian on X, collage vs riso b-roll skills (Oct 2026))
  2. Place each insert from the word-level transcript: in on the word that names the thing, out at the next breath. Few, precise inserts beat one every few seconds. Calliope's real-people b-roll pass pops each photo in the first time the narration says the name, with the example setting 'only historical figures, about 2 per minute' (306 bookmarks on the demo). (Sergio Gil on X, real people and places as b-roll (Oct 2026))
  3. Use what the person already owns first (screen recordings, product UI, old footage, photos); generate only the gaps. Nate Herk's open editing kit hard-codes the rule 'Use my existing footage before proposing generated assets'; real material also keeps the piece clear of YouTube's 'mass-produced' signal. (nateherkai/herk-edit-kit README (GitHub))
  4. Keep one look: write the style once (design-system file, palette, type, grain) and pass it to every shot; for generated clips name the shot size and one camera move and keep the light constant. Jason Zook's 3-step recipe starts by having Claude Code write a design-system.md from the app, then animates every interaction with it as YouTube b-roll: 240,406 views and 6,114 bookmarks (2.3 per like). (Jason Zook on X, Claude Code + Remotion b-roll from your design system (Apr 2026))
  5. Label realistic AI shots of real places or events (a small 'AI video' tag in the corner works) and keep a real face or voice in the piece. YouTube requires disclosure for 'AI generated extra footage of a real place'. In January 2026 it terminated 16 faceless channels (35M subscribers, 4.7B views) under its inauthentic-content policy and tuned recommendations toward real faces. (YouTube Help, disclosing altered or synthetic content)

Stopped working:

Prompts that work

Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}.

Route every line to the right b-roll (planner) (agent)

First step for any talking head or voiceover: decides where b-roll goes and what kind, before anything is generated or downloaded. Adapted from louisedesadeleer/b-roll-finder README (routing model, placement rules).

Here is my {VIDEO_TYPE} video: {TRANSCRIPT_OR_PATH}. Plan b-roll before sourcing or generating anything.
1. Transcribe at word level if needed. Read the whole thing and state the thesis in one sentence.
2. For each candidate beat write one sentence: what is this line actually about? Illustrate the point, not the nearest keyword. If nothing shows it accurately, drop the beat.
3. Route each beat: a named person, product, place or event: official footage or mine ({MY_ASSETS}); my product: my real UI as motion graphics; a number, order or comparison: a data card whose number matches the script; an opinion, feeling or metaphor: paper collage; a current claim: a clean screenshot of the source; something nobody filmed: a generated clip tagged "AI video".
4. Show me a table: timecode, quoted words, interpretation, route, source. Wait for my OK.
5. Place each approved clip full-bleed and silent, 0.2-0.5 s after the word that names it, out at the next breath. Dense in the hook, sparse after. Keep a b-roll manifest so approved beats never vanish between versions.
6. Grab a frame at every insert's midpoint and every joint, tile them, and check for wrong shots, burned-in captions or stray text before showing me.

Your UI as motion b-roll (design-system.md to Remotion) (agent)

Claude Code prompt that rebuilds your own product's interactions as clean 16:9 or 9:16 inserts in your brand, rendered locally for free. Adapted from Jason Zook on X, the 3-step design-system.md + Remotion b-roll recipe.

Step 1: Read this codebase ({APP_PATH}) and write design-system.md: colors as hex tokens, type families and sizes, spacing, radii, shadows, icon style, and the 5-8 core UI components with their states.
Step 2: Install Remotion in a new folder {BROLL_DIR}; do not touch the app.
Step 3: Using only design-system.md, make {N} motion-graphics clips of real interactions in the app: {INTERACTIONS: a card appears, a chat reply streams in, a notification lands, a checkout completes}. Each {3-5} s, {RATIO: 16:9 for YouTube}, 30 fps, plain {BG_COLOR} background, no voice, no captions, one clear action per clip, eased in and out, last frame held 0.5 s.
Optional: mock a familiar chat or email tool generically (no real logos) for context shots, and place my clip {CLIP_PATH} inside a frame at {TIMECODE}.
Render each to {BROLL_DIR}/out/<name>.mp4, then list the files with durations and the transcript line each one covers.

Collage b-roll: approved still, then assemble from empty (omni_flash_1_1)

One script line to a 5 s paper-collage insert. Send STILL to an image model (nano_banana_2_1 or gpt_image_2_5), approve it, then ANIMATION with the plain color field as first frame and the still as last frame. From pyang5166/gbro-collage-broll SKILL.md (imagegen and Omni prompt templates).

STILL (9:16): Create a finished editorial paper-collage image expressing {ONE_LINE_VISUAL_IDEA}. Perfectly flat {COLOR} paper field {HEX} with subtle uncoated paper fiber. Black-and-white halftone photo cut-outs with selective {ACCENT} cardstock. Central subject within the middle 70 percent, generous clean color field, 3-6 large separable paper groups. Visible halftone dots, crisp machine-cut edges, thin warm-cream keylines, soft low-opacity shadows. Must read at a glance: {RELATIONSHIP}. No typography, readable letters, numerals, logos, watermark, UI, subtitles, glossy 3D or photoreal environment.
ANIMATION: Paper-collage stop-motion assembly, using Image 1 as the exact empty first frame and Image 2 as the exact completed last frame. In one continuous locked-off vertical shot, open on the empty flat {COLOR} paper field. Assemble the scene piece by piece with crisp physical stop-motion timing: {STRUCTURE first, then SUBJECT, then CONNECTORS, then ACTION and RESULT}. End by holding the supplied completed composition. Preserve the exact 9:16 framing, {HEX} color field, cardstock accents, paper grain, halftone dots, cream keylines, cut edges and soft shadows. No scene cuts, no camera movement, no zoom, no morphing, no new objects, no text, no numbers, no logos, no sound.

Trending formats to remake

To remake one: download the example with the media-download skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below.

  1. UI motion b-roll from your own design system: The agent writes a design-system file from your app, then animates real interactions (cards, chat, notifications, flows) as clean 16:9 or 9:16 inserts for a talking head or launch video. Why it spreads: It shows the actual product, matches the brand and costs nothing to render. Jason Zook's recipe: 240,406 views, 6,114 bookmarks, the most-saved b-roll post of the year in this search. Keep: Keep a design-system.md written from your real app, one interaction per 3-5 s clip, a clean background and the app's own colors and type; swap the interactions and the app. Make it with Remotion Create (remotion-create).
  2. Editorial paper-collage b-roll: Still cutouts (paper, halftone, torn edges, stamps) generated per line without any text, animated in code, with all type set by the agent; ends on the first frame so it can loop. Why it spreads: Ideas and history have no footage; collage makes them concrete and reads as designed, not as AI stock. Alex Patrascu's printing-press piece: 638 bookmarks on 26,613 views (1.25 per like), full prompt shared. Keep: Keep stills with no text inside, all type and motion done in code, one flat color field per line, halftone cut-outs assembling from empty, every fact checked and the last frame equal to the first; swap the topic and the palette. Make it with Halftone Collage B-roll (gbro-collage-broll).
  3. Labelled AI cutaways over a real talking head: The speaker stays real; cutaways mix real clips with generated shots for what was never filmed (even the speaker at a desk), each tagged 'AI video' in a corner. Why it spreads: Keeps the human, fills the gaps, and the tag answers 'is this real?' before the comments do. Herk's reel with REAL CLIP / AI VIDEO tags: 505 bookmarks on 24,626 views; YouTube requires the disclosure for realistic shots of real places. Keep: Keep the real speaker, your own clips first and generated shots only for what was never filmed, each with a small AI VIDEO (or REAL CLIP) corner tag, cut in on the word that names it; swap the footage and the topic. Make it with varg AI (varg-ai).

Open-source repos

Set up varg

varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs.

Related tutorials