The AdFoundry System
One brand.json turns the video subscription you already pay for into a directed ad department.
A local command-line toolchain that fires your Higgsfield-generated shots from a written prompt doctrine instead of trial and error, tracks every attempt so you never re-spend credits on a repeat, and exports platform-ready cuts with one command.
One-time payment · instant download · unlimited use on your own machine
Runs on Python 3.11+, FFmpeg, and your own Higgsfield subscription · Claude Code optional
Real output, not a mockup
The doubt every buyer of an "AI ad tool" has: does this actually produce usable footage, or is it just a prompt template? Here's an unedited clip fired by higgsfield generate video kling3_0 — the exact CLI command family this package's video command wraps.
kling3_0, 5.04s, 3.76MB — fired live 2026-07-11 during our own Higgsfield provider integration testing, billed 10 credits on a standard Higgsfield subscription. Unedited, straight off the generation call.
It's not that AI video is bad. It's that you're directing it blind.
- You already pay for Seedance, Kling, or Veo credits every month — and still burn an hour hand-writing one-off prompts for every single shot.
- Every generation call starts from zero: re-explain the brand, re-describe the camera move, re-upload references, hope continuity holds between shots.
- Nothing remembers what you already tried, so a failed take gets re-generated with the same doomed prompt, on the same credits, again.
- Getting from raw clips to something postable costs another hour in a timeline editor — cutting, captioning, and re-exporting per platform by hand.
That's not a prompting problem. It's a production-discipline problem.
You're not buying a script. You're staffing six jobs.
Producer
adstudio initTakes your product, the claims you can actually prove, tone, and platforms — and locks a single brand.json before a single credit is spent.
Reference Wrangler
adstudio refsUploads once, hashes and caches every UUID, and enforces the eight-reference ceiling so a stray background never leaks into a shot.
Director
adstudio stills / videoFires batched generations built from a six-dimension prompt anatomy (who, action, camera, environment, light, sound) so every take reads like a directed shot.
Editor
adstudio judgeRuns an optional Claude-powered taste pass against a written rubric before you spend time reviewing footage by eye.
Post House
adstudio assembleCuts approved clips, burns captions, and exports 9:16 / 1:1 / 16:9 at -14 LUFS with one command.
Production Coordinator
adstudio trackKeeps an append-only ledger of every generation, so a re-run skips anything already completed instead of re-billing your credits.
How it works
Tell it what you're selling
Run adstudio init — answer a few questions about the product, the claims you can actually prove, the tone, and the platforms you're cutting for. It writes brand.json once.
Point it at references and let it direct the shoot
adstudio refs caches your uploads; adstudio video fires the batch from your request file. One line like "camera": "slow push-in from wide to medium, ending on the label" becomes a full directed prompt — you write the intent, the system writes the anatomy.
Approve and export
adstudio judge flags the weak takes, adstudio assemble cuts the approved clips into captioned, platform-ready exports.
A local ledger, not a spreadsheet ritual
Every fire is fingerprinted (model, prompt, refs, params) and appended to ledger/generations.jsonl before it runs. adstudio track reads it back before every batch, so a repeat request skips completed work instead of re-billing credits. Shape shown below — your real rows will differ.
| kind | model | status | output |
|---|---|---|---|
| video | kling3_0 | complete | generated/shot-01-push-in.mp4 |
| video | kling3_0 | skipped (ledger hit) | generated/shot-01-push-in.mp4 |
| still | nano_banana | complete | generated/shot-02-label-hero.png |
| video | seedance_2_0 | failed: ip_detected | — |
brand.jsonrefs/uuid-cache.jsonledger/generations.jsonlledger/generations.csvledger/judge-latest.jsonexports/ad-9x16.mp4exports/ad-1x1.mp4exports/ad-16x9.mp4What's actually inside
Six-dimension prompt anatomy
Every shot is directed, not guessed
Generation calls are built from WHO, ACTION, CAMERA, ENVIRONMENT, LIGHT, and SOUND so the model gets a complete instruction, not a one-line hope.
Reference hygiene, enforced
A stray background can't leak into your output
Uploads are cached by file hash, labeled by role and scope, and hard-capped at eight references — the ceiling Higgsfield itself enforces.
Content-filter map
Known rejections come with a rewrite, not a dead end
Documented nsfw / IP-detection triggers ship with pre-written safe rewrites, so a flagged generation doesn't cost you a redo cycle from scratch.
Append-only generation ledger
It compounds instead of repeating itself
Every fire is fingerprinted and logged before it runs. A repeat request against the same model, prompt, refs, and params skips straight to the cached result.
Optional Claude taste pass
A second opinion before you review by eye
When Claude Code is installed, `judge` scores your takes against a written rubric. No Claude Code on the machine? It skips gracefully — never required.
One-command platform export
Cut once, ship to every platform
FFmpeg concat, caption burn, and loudness-normalized (-14 LUFS) exports at 9:16, 1:1, and 16:9 in a single `assemble` call.
Made for the work you actually do
- Solo ecommerce founders who already have a Higgsfield or Seedance subscription and want the credits spent on directed shots, not trial and error.
- Agency editors carrying ad output for more than one client, who need a ledger instead of memory to know what's already been tried.
- UGC-style ad shops who need consistent references held across a whole campaign instead of restarting continuity every clip.
- DTC brand teams who need platform-ready exports without opening a timeline editor for every cut.
The AdFoundry System
$150 one-time
- ✓The adstudio CLI: init, refs, stills, video, judge, assemble, track
- ✓Six-dimension prompt anatomy + multi-shot timestamp grammar doctrine
- ✓Content-filter map with safe-rewrite guidance
- ✓Reference-hygiene rules (eight-reference ceiling, hash caching)
- ✓Append-only local ledger (JSONL + CSV) with read-before-fire dedup
- ✓FFmpeg assembly: caption burn, 9:16 / 1:1 / 16:9 exports at -14 LUFS
- ✓Worked example campaign + the 12-step manual
Delivered instantly to your inbox. All sales final — this is software, not a service or a subscription.
Questions before you buy? Email support@adfoundry.studio — we read every one.
Before you buy
Do I need to know how to code?
No. It's a command-line tool with a written manual — if you can follow numbered steps in a terminal, you can run it.
What do I need on my machine?
Python 3.11+, FFmpeg, and your own Higgsfield subscription. The system does not include or resell provider credits — it drives the account you already pay for.
Is this software or a service?
Software. It runs locally on your machine, using your own generation account. Nothing is generated on our servers, and we never see your references or footage.
Can I reuse it for future products?
Yes — unlimited local use, no recurring fee, no seat limit on your own machine.
What's actually in the download?
The adstudio CLI package, the full knowledge pack (manual, prompt frameworks, filter map, reference rules), and a worked example campaign request.
Stop hand-writing one-off prompts. Start directing.
One-time payment, instant download, runs on the generation subscription you already pay for.