This AI Agent Builds Full Video Ads For You (Claude Code + Seedance 2.0)

Mr. Paid Social · 53:55 · hook an AI agent up to a video-generation API so it clones winning ads from a single product image · watch

(Note: captions transcribe Seedance as “C Dance”/“SeaDance”, Veo as “Vio”, Arc Ads as “our cats” — corrected throughout below.)

The pitch

Drag any winning ad plus a product image into Claude Code; a free published repo (link in the video description) self-installs an agent that connects Claude Code to Arc Ads AI via its public API. Claude Code then writes the video prompt, fires off generation (Seedance 2.0 or others), retries failures, downloads results, and stitches clips into a finished ad — no chat copy-pasting between a custom GPT and Arc Ads anymore. All his demo products (“Mr. Paid Social”-branded Vans, tees, hats, cola, collagen) are fictional images made in Nano Banana 2.

Arc Ads basics

• Models available: Sora 2 / Sora 2 Pro, Kling / Kling 3.0, Veo 3.1, Seedance 1.5 / 2.0, Grok Video.
• Aspect 9:16 or 16:9; clip length 4–15 s.
• Credits scale with length (~10 s ≈ 0.6 credits) and halve at 480p.
• Counterintuitive tip: generate 480p deliberately for social/mobile feeds — lower resolution reads as authentic phone footage, less obviously AI.
• Any Arc Ads feature works through the agent once the API key is set: Seedance, Veo 3.1, Nano Banana Pro / Nano Banana 2 image gen included.

What Seedance 2.0 can take as references

• A prompt can mix multiple reference images + reference audio + reference video, selected inline with / references — e.g. “take reference video 1 as vibe, replace with product from image 1 using audio from audio 1.”
• He says it’s the first released model supporting both audio and video references together (Veo 3 only did images).
• Extend/continue: feed a generated clip back as a reference and continue past 15 s while keeping character, lighting, voice, even music vibe — used to stretch a Zillow-style listing walkthrough into what he estimates would cost ~$40 total vs. thousands for a videographer crew.
• Editing instructions work directly: “remove any visible production mistakes” removed a camera operator from frame.
• Start-frame + end-frame time-lapse is a killer use: give empty room → designed room (real estate), or woods → finished treehouse with interior walk-through.
• Green-screen output isn’t reliably green — his fix: generate the talking head normally, drop it in CapCut, one-click background removal, put the screenshot behind. Result looked fully real (moving hands, eye contact).

Ad styles demoed (mostly lazy one-line prompts)

• Influencer try-on (up to 9 reference images) of branded sneakers.
• POV ASMR unboxing: “no music, high mic gain, influencer whispering” — fingers and can detail held up.
• App promo built from a plain uncropped hero screenshot → scripted UGC bounce shots.
• Fast-cut Gen Z lifestyle collage; natural iPhone-quality lighting.
• Street-interview creatives (“two guys ask women what they think about Mr. Paid Social”) — an entire agency niche he thinks is now cloneable.
• Claymation/Pixar-style supplement narrative ads (cites Dara Denney’s breakdown of why these perform).
• Post-it-note reveal ad recreated from a Javy coffee ad found in the wild: rebrand screenshots in Nano Banana 2 first, then a start-frame prompt chain that peels post-its revealing each image. Handwriting reads slightly printed compared to the original.
• Voice cloning: record yourself saying anything, pass as reference audio — the AI character speaks your words, copies your laugh.

Setting up the Claude Code agent

1. Copy the GitHub repo URL (in description).
2. Create a local folder for the project.
3. In Claude Desktop (desktop app preferred over web/mobile — files land on your drive; terminal also fine), select the folder in a new session and paste the URL with “please set up this repo on my machine.” Repo contains its own setup instructions and skills.
4. In Arc Ads → Settings → Public API → get new credentials (client ID + client secret).
5. Paste the key into the repo’s .env file when Claude asks — not into the chat window (his security rule of thumb).

He runs bypass/autoaccept permissions mode for speed, with an explicit caveat that you’re trading safety for fewer approval clicks.

The clone-a-video skill workflow

This was added mid-recording to the repo and is the core trick:

1. Drop a reference ad (e.g. a Mott & Bow tee ad) plus a product image into Claude Code; say “clone this video for my product.”
2. Agent runs FFmpeg to extract frames (12 screenshots on one run), transcribes audio with Whisper, analyzes pacing — e.g. “39 seconds, 8 beats, 118 words,” beat map with narrative arc, framing, locations, tone (“I feel your pain → here’s the answer → you deserve this”).
3. Decides what transfers (hands-on-product moments, jump cuts, physical proof) vs. what changes (brand, claims).
4. Dialogue approval gate: presents script for review before spending credits; iterate conversationally if you don’t like it. He recommends giving more context than just an image — website URL, product description — for sharper dialogue.
5. Picks mode: 15-second compression, two-clip series (clip B fed clip A as reference), or both; shows estimated credit cost before firing.
6. Handles API pain automatically — one reference-image expiry triggered retries until success.
7. Downloads finished clips and stitches them into the final ad inside the project folder.
8. Say “commit this to a skill” to save any clone recipe as a reusable template — his library has feature-walkthrough, premium-reveal, product-hero, studio-look-back, UGC skills.

Known failure mode seen: small on-product text came out jumbled; suggested safeguard is generating starting-frame images (Nano Banana 2) and approving them before the full render — literally tell Claude to do this and it will.

His summary of the loop: “It’s basically vibe coding, but instead of vibe coding, you’re vibe creating.”

Live end-to-end example

• Source hunting via the Motion app’s Inspo tab: filter by visual format (testimonial etc.), days active (~3 months = likely winner); picked an anti-wrinkle face-mask ad.
• Cloned for fictional Collagen Peptides powder from a single product image: agent compressed 26 s → offered 15 s clip and/or two-clip series, did both, costed credits, generated in Arc Ads, downloaded, stitched a 30 s final. Verdict: nailed the UGC vibe (blemishes, couch, bathroom, real-looking iPhone); minor miss — background products all read “Mr. Paid” text, fixable with a counter reference image.
• Parallelism: open another Claude session on the same folder and clone different ads simultaneously — effectively unlimited concurrent agents.

Promo flags

• Private community “AI Ad Alchemist” on School (459 members at filming) pitched twice; bulk-ad system teased for members.
• Everything shown runs on paid Arc Ads credits beyond the free repo.