How to Add B-Roll to a Talking-Head Video With AI

What B-roll is, and which part AI takes over

Wikipedia defines B-roll as supplemental or alternative footage intercut with the main shot. In a talking-head video the main shot is you, speaking to the camera. The B-roll is everything else: your hands on a keyboard, the product you mention, the street outside, a dashboard on your screen. The same article notes that these secondary images are usually shown without their own sound, because the audio of the main footage carries on, and that in interviews and documentaries B-roll is often filmed after the interview to support what the subject said.

That order is the one this guide follows. You record yourself first, collect clips for the things you said, and then place them. The name itself comes from 16 mm film editing, when shots were split across two rolls in a checkerboard pattern to hide the splices.

Adding B-roll with AI here means handing the slow, exact parts to Claude Code: transcribing your speech, reading what each clip shows, lining clips up with sentences, and writing the ffmpeg commands that render the result. Two things stay with you: the footage and the final say on which clip goes where.

If you have never opened Claude Code for video, start with Claude Code video editing for beginners. It covers the install, the folder and the first render.

Step by step: adding B-roll with Claude Code

1. Put your clips in a broll/ folder

Make one folder per video, with the recording and a broll/ folder beside it:

my-video/
  input.mp4
  broll/
    keyboard-closeup.mp4
    coffee-pour.mp4
    analytics-dashboard.mov
    office-window.jpg

Name each file after what it shows. A descriptive name is the cheapest signal Claude has. It can also look at pictures: Claude Code’s documentation says you can give it an image path and ask what the image shows. For video clips, ask Claude to export one frame from the middle of each clip and write a one-line description of it into broll/index.md. Read that list once; a wrong description there becomes a wrong placement later.

2. Make a transcript with timestamps

Placement depends on knowing when each word is said. faster-whisper, a reimplementation of OpenAI’s Whisper built on CTranslate2, gives each word its own start and end when it is called with word_timestamps=True.

Transcribe input.mp4 with faster-whisper, word_timestamps=True.
Save words.json (each word with start and end in seconds) and
sentences.json (each sentence with its first word's start and its
last word's end). First fix these spellings: [NAMES AND TERMS].

Check the names of products and people before going further. Claude matches clips to words, so a misheard brand name means a missed match.

3. Let Claude match clips to sentences and write a plan

Switch Claude Code to plan mode with Shift+Tab. According to the Claude Code docs, in plan mode Claude reads files and proposes a plan but makes no edits until you approve. That is the right mode for this step, because nothing should render yet.

Read sentences.json and broll/index.md. Plan B-roll for input.mp4.
Write broll_plan.md as a table: start, end, the sentence, clip file,
in-point inside the clip, full frame or corner, and one line on why.
Rules: start and end on sentence boundaries; no slot over [4] s;
no clip used twice; keep the first [3] s and the last sentence on me;
leave out sentences where nothing in broll/ fits. Do not render.

This is different from the short B-roll notes prompt in AI video editing prompts for Claude Code. That prompt lists moments where something could be shown, before you have footage. This one maps the clips already sitting in broll/ to exact times.

4. Check the plan

Read broll_plan.md with the video open. Look for four things:

• Meaning, not keywords. “We burned the old plan” should not get a clip of a fire.
• Your key moments. A sentence where you point at the camera, react or show something in your hands should stay on you.
• Rhythm. Three cutaways in a row with no return to your face feel like a slideshow.
• In-points. A clip that starts on a camera wobble needs a later in-point.

Change the table directly, or tell Claude what to change. Approve the plan only when you would be happy to see it on screen.

5. Render with your voice kept under the B-roll

ffmpeg’s overlay filter takes two inputs and has one output; the first input is the main video and the second is laid on top of it at the x and y position you set. The documentation recommends passing both inputs through setpts=PTS-STARTPTS so they start at the same zero timestamp. One default matters here: when the overlaid input ends, eof_action repeats its last frame. A clip shorter than its slot therefore freezes on screen, so every clip should be trimmed to the exact length of its slot.

Render broll.mp4 from input.mp4 and broll_plan.md.
- Audio: from input.mp4 only, unchanged. Mute every B-roll clip.
- Trim each clip to its slot from its in-point; scale and crop it to
  fill the frame, or to a [30]% corner box for rows marked "corner".
- Keep the size, fps and length of input.mp4.
First render a 540-wide preview of the first two slots and wait.

6. Export and watch it once

After the preview looks right, ask for the full render and export it at the size you publish. Then watch it from start to finish with sound. Listen for gaps or doubled audio at each cutaway, and look for frozen frames at the end of a slot. Claude can check still frames for you, but timing and feel are easier to judge by watching.

How to shoot or choose B-roll that matches what you say

The best B-roll is planned from the words, not collected at random. Before you shoot, give Claude your transcript and ask for a shot list: every sentence that names a thing, a place, an action or a result, with one idea for what to show.

Wikipedia lists insert shots and establishing shots among the kinds of B-roll, and both suit talking-head videos. An insert is a close detail: a hand signing a form, a cursor clicking a button. An establishing shot shows where you are: the outside of a shop, a desk from across the room. A mix of the two gives the viewer a reason to look at each cutaway.

A few habits make the footage easier to place:

• Shoot longer than you need. Hold each shot steady for several seconds so there is room for an in-point.
• Match the orientation. Film vertical B-roll for a vertical video.
• Show the literal thing first. If you say “our dashboard”, show the dashboard, not a stock clip of a laptop.
• One idea per clip. A clip that pans across three things is hard to match to one sentence.

Placement rules that keep the edit watchable

These rules are worth writing into every plan prompt.

Keep the audio continuous. Your voice runs from the first frame to the last; only the picture changes. That is how B-roll has traditionally been cut, with sound from the main footage.

Cut on sentence boundaries. A clip that starts halfway through a word looks like a mistake. Word end times from the transcript give clean in and out points.

Use B-roll to hide your own cuts. Wikipedia’s article on cutaways says they are used to conceal the deletion of an unwanted part of the main shot, and that in interviews cutaways to the interviewer can cover cuts. When you trim a stumble or a filler, a short clip over the join hides the jump.

Come back to your face. A cutaway is usually followed by a return to the first shot. Viewers of a talking-head video came for a person; long stretches without you lose that.

Use a corner when both matter. If you are explaining something on screen while reacting to it, a corner box keeps the picture visible without taking you away.

No footage? Use stills, screen recordings or images

Many talking-head videos do not need filmed B-roll at all. Screen recordings work for anything you do on a computer: record the steps you describe, then let Claude place them the same way as camera clips. Photos and screenshots can sit in broll/ too; ask Claude to hold each one for its slot, with a slow zoom if a still feels too static.

Two styles are built around images instead of footage. Cut-out Backdrop cuts you out of your room and places you over full-screen images that change with every idea; you bring one image per idea, at least 1080 px wide. Image Cluster keeps you full frame and floats small groups of two to four images above your head, and Claude can propose a list of what to collect.

Rights to stock footage, photos and screenshots

Stock footage is the usual fallback when you cannot film something, and its licence matters. Stock suppliers may offer footage on rights-managed or royalty-free terms. Royalty-free means you do not pay a fee for each use; it does not mean free of cost, and the work is still under copyright.

Creative Commons licences add conditions you need to read. Attribution (BY) requires a credit to the author. NonCommercial (NC) allows use only for non-commercial purposes, so check whether a monetised or sponsored video counts. NoDerivatives (ND) does not allow sharing adapted versions. CC0 is a tool for waiving as many rights as legally possible. Wikipedia also notes that the United States government is one of the largest producers of public domain stock footage.

Keep a credits.txt in the folder with the source and licence of every clip you did not shoot yourself. Claude can turn it into the credit lines for your description. This is a practical checklist, not legal advice; when a licence is unclear, ask the rights holder.

Ready-made styles built around your own B-roll

If you add B-roll to every video, a skill saves writing the plan prompt each time. B-roll and Stacked Text cuts pauses and filler, covers the video with B-roll every few seconds and stacks the key phrase of each sentence on screen; you point it at your clips with “B-roll is in broll/”, and it does not shoot or source the clips itself. Personal Reels cuts on spoken words every 0.6 to 1.6 seconds between framings and your own B-roll, and asks for 5 to 15 short clips or photos, such as your desk, hands, screen, street or product, before it edits. Both export a 1080×1920 video and run on macOS or Linux with Python 3.9+ and ffmpeg.