Written by Artufrfr (NO AI)

After reading this post you will learn:

- How to create AI podcast-style Video Ads using Seedance 2.0/Higgsfield

- How to use Claude to turn a Script into a Realistic Scene Prompt

- How to use the Free SeeDance 2.0 Prompt

- How to Keep Character Consistency across Video Generations

- How to Edit AI Clips into a Polished Ad in CapCut for High Conversion

Seedance 2.0, released on February 12, 2026 & currently the best text-to-video model for generating realistic human conversations on camera.

A two-person podcast-style footage with product interaction and integrated audio.

The video at the top is an example of what the finished video looks like. It feels like any other podcast clippers create clips for. The organic feel is what gets your viewer to stop thinking your posts are just bad ai slop and it's what gets even the most tech-savvy audience to buy your products.

And yours can be better, for example, if I specified the name of the podcast it’d look more real rather than just a LED sign that says ‘talk’. Other than that everything else looks super real. The average viewer would assume the characters & the podcast are all legit.

The reason this format works for ads is the same reason the best podcast clips go viral. It doesn't feel like an ad. Read this post on advertorials if you want to know how to format your ads in a way that doesn’t scream “buy my product link in bio”.

This is what the workflow looks like:

Claude (Script) -> Seedance 2.0/Higgsfield (Video Generation) -> CapCut (Editing + Captions + B-roll)

Level 1: Writing Your Script with Claude

cost: $0

We’re going to start with the script. Remember, It isn’t what your characters say, but how they say it, mini-movements, pauses & natural body language makes the Seedance Avatars seem like real people and not robots.

This is where most people go wrong. They write a dialogue and paste it straight into a video generator. The output looks robotic because the prompts are robotic & Claude fixes this.

Step 1.1: Define Your Two Characters

Before you write a single line of dialogue, define your host and guest clearly in the prompt detailed in 1.3, just replace the placeholder of description of the host and guest. Include age, appearance, clothing & energy. Be specific as Seedance will build the avatar directly from this description.

For example in the video I used:

Host: 30-year-old white male, nerdy with glasses, wearing a backwards cap, casual podcast energy.

Guest: 50-year-old Black male, visibly muscular, wearing a USA military uniform, calm and authoritative.

You will paste this within the Claude prompt before it goes into Seedance.

Step 1.2: Write a Rough Dialogue

Write a basic back-and-forth script in the same prompt in 1.3’s intro.

Host: So I heard you took supplements in the military (question)

Guest: Yeah we took [product] (answer)

Host: Oh wow [reaction]

Guest: Yes so we took [supplement] for 5 days

Clip 2/3

Host: Question

Guest: Answer…

Host: omg so cool let me syd…

Host: I’m going to start taking it too

Don't overdo the first draft. Again, you're going to hand it to Claude to give it a Director’s touch.

Step 1.3: Send It to Claude to Refine

Now paste your rough dialogue into Claude AFTER sending the following prompt:

"Imagine you are a director writing a podcast scene for social media, shot in a modern, slightly moody professional studio and edited into 9:16 short clips of the highlights. The setup uses a 3-camera multi-angle system. Cam A is a medium-wide two-shot, Cam B is a close-up on the host, Cam C is a close-up on the guest, with subtle jump cuts between angles every 2 to 4 seconds and occasional barely-noticeable digital zoom-ins. Framing is slightly off-center at times, imperfect and naturalistic. Lighting is warm with soft shadows, mild contrast between subjects and background, and visible practical lights (LED panels, neon sign, desk lamp) with minor inconsistencies across the frame. The background features a lived-in podcast desk setup, boom-arm mics, headphones, acoustic panels, shelves, plants, and a neon sign, with shallow depth of field giving a slight background blur. All shots are social-first: vertical 9:16 or center-cropped, chest-up, with natural headroom and asymmetrical composition. Audio is clean podcast-mic quality but not over-processed, slight room tone, natural breaths, mouth sounds, and mic proximity variation are all present. A product (provided via image reference) is held casually by the guest throughout, loosely in one hand near chest or table level, with a relaxed and slightly imperfect grip that shifts naturally over time. The product tilts, rotates slightly, and drifts in and out of frame with hand gestures. It is never centered, zoomed into, or treated as a focal point, it simply exists in the scene. The host may briefly glance at it but never fixates. Both characters display continuous natural behavior: subtle head nods while listening, micro-expressions, natural non-synchronized blinking, slight posture shifts, small hand gestures near chest or table level, audible breathing between lines, occasional lip presses and mid-thought pauses, and 100–400ms delays before speaking. The host behaves as an active listener, reacting with varied short affirmations, occasionally interrupting lightly, leaning forward with visible curiosity, and responding differently each time with semi-random, unscripted energy. The overall clip runs approximately 15 seconds, fast-moving but natural, with each line lasting 1–2.5 seconds, slight overlaps, and micro-pauses that make it feel like a clipped highlight from a longer episode. Cinematically: very subtle handheld micro-movements, minor focus breathing on close-ups, tiny exposure shifts between cuts, no over-sharpened or AI-smoothed image quality, natural skin texture, and subtle background ambient motion. Take this dialogue and rewrite it with detailed, natural stage directions for each line. Include small nods, a breath in before speaking, subtle hand gestures, eye contact shifts, and product handling moments. Make the body language feel completely human and unscripted. Here are the character descriptions: [paste host and guest descriptions]. Here is the dialogue: [paste your script. What I want you to do is output: the Full refined Seedance 2.0 prompt, keep the same structure and sections, integrate all improvements directly into the prompt, Don’t explain your changes, Don’t Summarize, Don’t remove user-editable placeholders.”

Claude will output a fleshed-out scene prompt, the dialogue will include descriptions of how each person physically moves and reacts moment to moment.

1 thing to note: Claude will sometimes spit out a character description summary at the top of its output. Ignore that part and copy everything below it, that's your actual scene prompt.

Step 1.4: Keep Each Clip to Four Lines Maximum

Seedance generates in 15-second windows. Too much dialogue in one clip and the avatars speak unnaturally fast or the clip cuts off mid-sentence. The sweet spot is four lines per clip, roughly one sentence each, two from the host and two from the guest but you can change this around how you like.

A complete 40-second ad is typically three clips. So it’s best to plan your script in blocks of four lines.

Level 2: Creating the AI podcastVideo Ads using Seedance 2.0/Higgsfield

Step 2.1: Change Settings in Seedance 2.0

Once your prompt is filled out, open Higgsfield and go to the video section. Select Seedance 2.0 and paste the full prompt in, then set:

Sound: On

Duration: 15 seconds

Aspect ratio: 3x4 for Instagram/Facebook (or 9x16 for TikTok)

Resolution: 1080p

Now hit generate. This first clip is important as it is your reference.

Step 2.2: Maintain Character Consistency for last 2 clips

If you only paste the prompt with new dialogue for your second clip, Seedance will generate slightly different-looking avatars even with the same character description. To ensure consistency, upload screenshots from your first clip alongside every subsequent generation:

A screenshot of both characters together in the studio

A screenshot of the guest alone

A screenshot of the host alone

A clean image of your product (white background works best)

Now with those four references uploaded alongside the prompt, Seedance produces the same faces/studio/voices, and adds new dialogue and product interaction.

Step 2.3: Generate Your Product Reveal Clip

Your second clip is where the guest pulls the product out. Write this scene in Claude the same way, describe the guest reaching for the product, placing it on the table, the host leaning in. Claude will add the body language detail.

If Claude accidentally swaps the host and guest roles (it happens occasionally), just tell it: "Edit the dialogue so [character A] is the guest and [character B] is the host." Fixed in seconds.

Step 2.4: Generate Your Closing Clip

Third clip is the close. Guest explains the main benefit simply. Host says they want to try it. Same process where you add four lines max, refine the script with claude, reference screenshots uploaded. And that is all your raw ad clips done.

Level 3: Editing in CapCut (Subtitles + B-roll)

cost: $0-$5/mo (CapCut Pro via G2G)

Your raw Seedance clips will have awkward pauses, filler sounds, and transitions that don't quite flow right so post-production editing fixes all of that. A full ad edit takes under 10 minutes, with b-roll it won’t take over 30 minutes maybe after the first or second time.

Step 3.1: Import and Rough Cut

Drop all three SeeDance clips onto the CapCut timeline. Watch through and cut anything that doesn't serve the conversation, filler sounds, long pauses between lines, any moment where the movement or audio feels slightly off.

Natural audio bleed is another trick for making AI podcast footage feel real. So get the audio from the next clip and then let it overlap slightly with the tail of the previous one. This makes it sound like one continuous conversation rather than three separate takes stitched together.

Step 3.2: Add Colour-Coded Captions

Add auto-captions to generate subtitles for the full video. Colour-code them by speaker, blue for the host and orange for the guest, or whatever combination fits your brand. It makes the video look intentional and helps viewers track the conversation without having to think about it.

Add a shadow effect to all captions so they're readable over any background. Position them in the middle of the frame.

Step 3.3: Add Background Music

Drop in a subtle background track. For a supplement or health product, something with low curiosity-building energy works well, understated but present. The music should sit underneath the dialogue, not compete with it. Look at tiktok reels for inspiration.

Step 3.4: Add Keyframe Zooms for Retention

At the 3-5 second mark of your first clip, add a slow keyframe zoom. Set a keyframe at the start (no zoom), then another near the end of that section with a slight push in, around 5-10%. This signals to the viewer's brain that something is building, which keeps them watching past the drop-off point.

Repeat on any moment where a character makes a strong statement or reacts visibly.

Step 3.5: Add B-roll

Cutting in 1-3 seconds of B-roll between your podcast clips makes the whole thing more engaging and less like a static talking-head video. Search YouTube Shorts for relevant footage, "muscle building," "military training," "gym", extract short clips and drop them in at natural transition points.

Add a swoop sound effect on each B-roll cut. Small detail, noticeable difference in how the energy feels.

Level 4: Test constantly.

The podcast format earns trust because it doesn't look like an ad and is common with alot of the clips currently on Tiktok, Instagram & Facebook. The dialogue feels genuine. Watch the finished video back with fresh eyes the next morning before you post it. If anything sounds rehearsed, salesy, or misplaced you should go back and fix it.

Test different hooks in the opening line. Test with and without B-roll. Test different products reveal timings. Change one variable at a time so you know exactly what's working. Scripts that convert well can be reused across completely different characters and products, it takes minutes to regenerate the whole thing with a new avatar once you have the template dialled in.

Congratulations, you just made a fake Podcast Interview clip and it looks just like the real thing! No need to pay for a whole podcast episode, no podcast equipment or fancy camera.

A well-made podcast-style ad can run as a first post on an account with 0 warm up on Facebook, Instagram, or TikTok with zero brand recognition and still convert, because the format builds trust in the viewer’s mind way before the product is even visible. That's the edge here that’s better than Heygen.

The quality of your Claude prompts determines the quality of your footage. Seedance is only as good as the scene description you feed it. Spend the most time on Level 1, tinker with the Claude prompt in Level 2 slightly to get the results you want , and the rest is just editing.

I hope this helped. If you’ve got any questions let me know and I’ll get back to you. Thank you for reading! If you need additional resources or help let me know you can find me anywhere.