How to Prompt B-Roll that Matches a Talking-Head Cut

Anyone sitting on a locked talking-head edit can prompt B-roll that matches a talking-head cut in one session: one coverage card per hole, one short clip per card, then stop before you recut it. Run it in Runway, Kling, or Grok Imagine. The talking head is the lock.

Pretty B-roll that ignores the line is stock. Generative clips arrive overlit and too long, then you recut the A-roll to hide the miss.

If you still paste "b-roll of a founder walking," read what a skill prompt actually is and save the card at the end.

Match B-roll to a talking-head cut in five steps

1. Lock the talking-head cut first

Do not generate from the topic. Generate from the timeline. For every hole, write four facts:

• Timecode in and out.
• Duration to the tenth of a second.
• The spoken line the insert must survive.
• What the viewer should see so that line is still true.

Example: "01:12.4–01:15.1, 2.7s. Line: we still roast on Tuesdays. See: one bag on a steel table, no café."

If you cannot name the line, you have a wish. For assemble, grade, and export, use the video prompt to finished-edit workflow. This skill is only coverage.

2. Freeze the look of the A-roll

Matching B-roll shares a room, not a brand adjective. Freeze a look you will not edit:

"Handheld-adjacent, 35mm feel, overcast window, no haze, no flare, no drone, no slow-mo."

Put that paragraph at the top of every prompt. If the talking head is a messy kitchen, do not prompt a marble island.

In Runway, Kling, or Grok Imagine, attach one still from the interview. Text freeze plus a still. A folder of vibes is how the clip leaves the building.

3. Prompt one insert at a time

One card, one prompt, one clip. Never a pack of options. Three lines under the freeze:

• Subject — counted. "One paper bag, closed, on a steel table."
• Motion — one verb. "Static" or "hands enter from frame left and leave."
• Duration and rejects — "3 seconds. No logo, no face, no extra props."

Tell the model: "Do not add a second action. Do not beautify." Models add steam and a second bag "for composition." Reject those frames.

4. Generate short, then trim on the timeline

Generate two to four seconds. A 10-second hero is a different job. Pull the clip onto the locked sequence. Trim to the line. If motion peaks after the line, the clip is wrong. Do not slide the interview.

Reject on any of these:

• Eyeline or horizon that fights the A-roll.
• Motion energy louder than the voice.
• A color that only exists in the insert.

Do not grade your way out of a reject. Store the freeze and the reject list in the Promptcrates library of reusable prompts.

5. Stop before you recut the A-roll

Inserts now fit holes you already decided. Do not recut the talking head to save a prettier miss. Do not generate a matching set of five. Do not swap from Kling to Runway mid-edit.

If three cards failed, the freeze was wrong. Start a new loop. Save this one with the Promptcrates tools for video skills.

Reusable skill card

Trigger: A locked talking-head cut, and you are about to type a topic into Runway, Kling, or Grok Imagine instead of a hole.

Input: Cut list (timecode, duration, spoken line, picture), freeze paragraph, one interview still.

Output: One 2–4 second clip per card, freeze unchanged, one motion verb, trimmed to the hole.

Stop: After inserts sit on the locked timeline. No A-roll recut. No pack. No model swap. No grade in the generator.

Takeaways

• Prompt from the cut, not the topic.
• Freeze the A-roll look. Verbatim tokens every insert.
• One card, one clip. Packs are stock.
• Generate short. Never slide the interview to save a miss.
• Save the freeze as a skill. The last fountain shot is disposable.