No More Higgsfield — Ad Creative Engine

by Mike Futia — https://www.skool.com/scale-ai/about

A Claude Code skill that replaces subscription AI creative tools (Higgsfield, Krea, Freepik AI) with pay-per-asset generation through a single fal.ai API key.

Drop in a winning Meta static ad and get back a complete testing matrix: copy variations (different hooks, angles, CTAs), new static ad images via GPT Image 2, and the winners animated into video via Kling or Seedance — all grounded in your brand’s actual guidelines, voice, and customer language.

A 5-variation testing cycle costs about $0.35. Animating a winner adds ~$0.50. Slow months cost $0.

Why this beats a subscription

• Pay for output, not capacity. No $49–$99/month whether you generate or not.
• It reads your brand. Approved claims, banned words, your customer’s exact language — every generation is grounded in your brand files. No subscription tool can do this.
• Cost quoted before anything fires. Every run shows a per-asset cost table first, and a hard budget cap is enforced in code — the script refuses to submit anything over budget.
• You own everything. Every asset downloads to your disk immediately, every prompt and cost is logged, and a local gallery shows your whole library. No deletion windows, no content-rights fine print.

Install

Claude desktop app (recommended): download ad-creative-engine.skill from this repo, then in the app go to Customize → Skills → Upload skill and select the file. Done.

Claude Code CLI: clone straight into your skills folder:

git clone https://github.com/mikefutia/no-more-higgsfield.git ~/.claude/skills/ad-creative-engine

Both paths then need the Python environment (one-time — or just ask Claude: “set up the ad-creative-engine skill’s Python environment” and it’ll run this for you):

python3 -m venv ~/.config/ad-creative-engine/venv
~/.config/ad-creative-engine/venv/bin/pip install fal-client python-dotenv requests

Finally, get a fal.ai key (fal.ai → Settings → API Keys), load $10 of pay-as-you-go credit, and save it:

mkdir -p ~/.config/fal && echo 'FAL_KEY=your-key-here' > ~/.config/fal/.env && chmod 600 ~/.config/fal/.env

(A .env file in your project folder or an exported FAL_KEY env var also works.)

Use

In a project folder, set up your brand context once — Claude drafts it for you:

Draft my three brand foundation files from https://yourbrand.com — brand/brand-dna.md
(products, approved claims, visual identity, what we can NEVER say in Meta ads),
brand/brand-voice.md (tone, do/don't words), and brand/icp-cards.md (target customer,
pain points, voice-of-customer phrases, objections).

Review the drafts — every future ad inherits them. Then drop your winning ad image into the folder and run the matrix:

Use the ad-creative-engine skill. My winning static ad is reference-ad.png and my brand
files are in brand/. Generate 5 ad variations spanning at least
3 different ad formats — each with a different hook angle, plus full Meta ad copy for each. Keep the total under $3 and
show me the cost quote before you generate anything.

Pick your winners, then:

Use the ad-creative-engine skill. Animate variation-3.png into a 5-second ad video with
Kling — slow push-in, subtle atmosphere. Quote me the cost first.

And to browse everything you’ve ever generated:

Rebuild my generations gallery and open it.

What’s in the box

Guardrails baked in

1. Cost table shown before any generation
2. Hard budget cap enforced in code (default $3/session) — over-budget manifests refuse to submit
3. Batch brake: >12 jobs or >$5 in one run requires explicit confirmation
4. Every output saved locally immediately; every prompt, cost, and seed logged to generations/log.jsonl
5. Never generates ad copy without your brand foundation files

Notes

• Prices drift. The skill verifies a model’s current fal.ai price on first use and updates its own price table.
• Supplement/health brands: the brand-dna template includes Meta compliance restrictions (no second-person health states, no outcome promises). Always run your own policy check before launching.

---

Built with Claude Code. From the SCALE AI community.