Making a 41-second car ad in one afternoon with the open-source Ad Storyboarder: Opus 5.5 in Claude Code directs, Higgsfield makes the images and video, and a read-only inspector looks for faults
Chinese creator Feicai (@zhu185178) made a 41-second self-initiated concept ad for the new Tesla Roadster. He wrote the creative script, answered a few questions, said once that "the car looks fake", and reviewed the cut; Opus 5.5 in Claude Code did the rest, generating art assets and 34 video clips through the Higgsfield CLI, using about 1,750 credits across two cuts. He released the workflow as an MIT-licensed .claude folder with six skills, a read-only inspector agent, measure-only check tools and a frame-by-frame review Studio. We did not run Higgsfield generation; we ran the free self-check, contact-sheet, loudness and delivery-gate tools on our test machine.
Visit site 广告分镜师 Ad Storyboarder (Claude Code + Opus 5.5 + Higgsfield CLI) ↗- Pricing
- Paid
- Region
- Global
- Where it's from
- Chinese community
- Added
- Oct 22, 2026
- Tested on
- Oct 22, 2026
- Version we used
- feicaiclub/ad-storyboarder main @ ea42037 (2026-10-19); local tools run on Debian Linux with ffmpeg 7.1.5, Python 3.13.5, numpy 2.5.4, Pillow 12.3.0
- Original author
- 废才 Feicai(废才俱乐部) @zhu185178 ↗
- Our test
- Partly tested
- Cost / time
- The workflow is free (MIT). Generation is billed in Higgsfield credits: the author is on the Ultra plan and this ad took about 1,750 credits across two cuts. You also need Claude Code (he only tested with Opus 5.5). His 2026-10-15 notes put Seedance 2.5 at 3 / 7 / 16 credits per second for 480p / 720p / 1080p; prices change quickly, so check with higgsfield generate cost.
The task
You have product shots and an idea, and want AI to make a commercial or pitch sample where the product and people stay accurate, nothing breaks and spending is controlled, without hand-writing every shot prompt and picking and editing clips yourself.
Steps
- Install Higgsfield: on Higgsfield's MCP page pick the Claude Code tab and paste the official install prompt into Claude Code, or by hand: npm i -g @higgsfield/cli → higgsfield auth login → higgsfield workspace set <workspace> (without it every call fails with No workspace selected) → npx skills add higgsfield-ai/skills.
- git clone https://github.com/feicaiclub/ad-storyboarder.git my-first-ad, cd in and run claude; for an existing project, copy the .claude folder in. You need ffmpeg, Python 3 with numpy and Pillow, and a CJK font; local whisper if there's dialogue or voice-over.
- Before starting it runs a self-check (python3 .claude/tools/doctor.py): Higgsfield login, workspace, credit balance and missing local tools.
- Hand over product shots, your idea and requirements, and the shooting script; if someone appears on camera, add photos of that person (yourself or with consent). Without a script it first interviews you, offers several distinct creative directions, then writes the script. The author's script was written beforehand with another model, down to each shot.
- Answer its questions: target screen, resolution, credit budget cap, acceptance criteria, whether to add an AI-generated label. The author didn't have enough credits for 1080p and chose 720p, 16:9.
- It splits the script into a segment table (six columns per segment: script, references, generated, reviewed, in the cut, accepted), writes a product spec sheet for colors, logo, materials and structure, prepares art assets, and test-shoots one segment at 480p before generating the rest.
- Before a segment goes into the cut, a read-only inspector (agents/inspector.md) pulls frames and checks ten things including product, faces, breakage, AI look and sound, returning a timecoded issue list; usable stretches are marked to the second and only those are edited in.
- The final cut must pass the gate (gate.py: A/V sync, silent gaps, clipping, loudness, true peak, no .env or key-like filenames in final/); a failure exits with code 2 and blocks delivery.
- Type /ad-studio to open the Studio: space to play, arrow keys to step frames, M to write a note, saved with a frame grab and timecode to 反馈/待处理.md (feedback/pending); hand over all notes at once, then use Compare to watch cut 2 beside cut 1.
Prompt
Here are my product shots, creative brief and shooting script. Make a 16:9, 720p ad using the Ad Storyboarder workflow. Run the self-check and report my credit balance first, and only ask questions the script can't answer. Test-shoot each segment at 480p first, and ask me before spending beyond the budget.
What the author says
In his 2026-10-19 article and the README, the author says he handed over the creative script at 12:05, finished answering questions at 12:19, said once at 13:10 that the car looked fake (it had been misled by official renders), got cut 1 at 15:30 and cut 2 at 17:37, and mostly left it alone. It generated 34 video clips, with about 16,000 English words of prompts all written per shot by the agent; Higgsfield credits went from 1,753 to 1.82. Before cut 1 the inspector found 6 must-fix issues, such as the glass canopy turning black in the barrel-roll shot and a different car bursting out of the white smoke. He says that with an idea, an ad like this can be made in an afternoon.
Our test
We have no Higgsfield account or credits and did not run the full workflow in Claude Code, so generation quality, the 41-second ad, credit use and the inspector's effectiveness are unverified. We only ran the model-free, no-cost tools on a Linux test machine: self-check, contact sheet, per-second loudness, delivery gate and the credential-guard hook, using test videos synthesized with ffmpeg.
What we found when we tried it:Tested 2026-10-22 (repo main @ ea42037, Debian Linux, ffmpeg 7.1.5, Python 3.13.5, numpy 2.5.4, Pillow 12.3.0): doctor.py --no-hf found ffmpeg, Python, numpy, Pillow and a Noto CJK font, flagged that local whisper was missing, and ended with a pass. The README says it was only run on macOS, but these local tools worked on Linux. qa.py sheet turned a 6-second 720p test clip into a 2400×944 contact sheet; qa.py levels reported -21.8 LUFS and no silent gaps for a 440 Hz sine track, and a silent gap (0, 5) for a silent clip. gate.py with a -14 LUFS target failed on loudness (-21.8 vs -14.0±1) with exit code 2; with a -22 target it passed with exit code 0. secrets-guard.sh blocked higgsfield auth token and a command reading ~/.config/higgsfield (exit 2) and let a plain ls through. We did not test generation, assembly, the inspector or the Studio page.
How to get in
The article is an X Article in Chinese (login required); the GitHub repo has a full English README (README.en.md). Higgsfield and Claude Code are available internationally. The tutorial video is on the author's Bilibili and YouTube (@feicaiclub) channels.
Who it suits
Ad makers, e-commerce video creators and small teams who need pitch samples; anyone who wants to see how an "agent as director + inspector + gate" harness is built. Not for people expecting the AI to supply the idea: the author is explicit that the idea must be yours.
Pitfalls
- No model does everything: the author's biggest regret is that in full-auto mode most art assets came from GPT Image 2.5, which got the shapes and his face most accurate but made skin and paint look greasy and dirty. He now uses Seedream, Nano Banana Pro or Midjourney for portraits and Seedance 2.5 for video, and says to fix the model choice before starting.
- Models copy whatever the reference shows: official studio renders as references gave the car a CG look, and the studio floor and camera angle can be pulled straight into the shot. Keep references clean with simple backgrounds, and say in the prompt to copy only the look and shape.
- Draft upgrades are billed at 1080p: in his tests a draft "upgraded to 720p" was charged at the 1080p rate and came out at 1080p. For a 720p delivery, generate directly at 720p except for the test segment.
- Avoid keyframes as first frames: in cut 1 the close-up after the driver steps out started from a separately generated image and his face changed. He switched to character and product references plus a scene image.
- Physically complex moves deform (the car roof bulges in the second after the barrel roll); he hid it with a cut to a reaction shot instead of re-rolling.
- Only tested on macOS + Claude Code + Opus 5.5. You still need the eye to spot a fake-looking car or a continuity break; the author caught cut 1's break himself.
- Real people need consent; the Roadster design belongs to Tesla and is demo-only, not covered by the MIT license. Higgsfield can block jobs on IP detection (the author logged an ip_detected status, with credits refunded).
More in this category
- A demo video per docs page that flags itself when the UI changesVideoScreenplayClaude CodeTool free; uses your agent's quota◐ Partly testedOverseas
- Code-drawn stick-figure explainers for about $0.54 eachVideoClaude CodeFFmpegChromeNeeds Claude plan or API; renders locally○ Not tested by usChina
- One topic in, a narrated tech explainer video out, locallyVideoClaude CodeChatterboxffmpegSkill free; needs a Claude Code plan◐ Partly testedOverseas
- Render restrained motion graphics from code, frame by frameVideoClaude CodePlaywrightffmpegNeeds a Claude plan; other tools free◐ Partly testedOverseas
- One reference image → a take-apart machine simVideoteardownClaude Codethree.jsSkill free; needs Claude Code/Opus◐ Partly testedOverseas