/generate

One command for media. Layer 2 (you) picks the model and enforces the money rules; Layer 3 (execution/generate_media.py) does the API call, the polling, the download, and the log. Never hand-roll an API call for this. If a model is missing, add a row to MODELS in the script and a recipe in models/.

Models

Read the recipe file before the first run of any model in a session. It has the request shape, the response shape, and the gotchas.

What is actually funded right now

python3 execution/generate_media.py list is the live source of truth.

1min.ai is the lane to prefer. The credits are already bought, so a draft there costs nothing at the margin while nano-banana costs real cents. Measured rates: 9,000 credits a schnell image, 90,000 a flux-2-pro image, against an 18.9M balance — about 2,100 or 210 images. It has no reference-image support, so --ref auto-switches to Google and says so.

Video is the one lane still on real money (Veo 3.1 is $0.80 to $4.80 for an 8s clip, depending on Fast/Standard and resolution; Higgsfield kling $1.12 for 10s, seedance-2.5 $9.10 for 8s at 1080p, quoted at list). 1min.ai hosts Veo3/Kling/Sora on prepaid credits and wiring that is the biggest saving left on the table — see the gotchas in models/onemin.md.

Provider routing

1. Default to the lowest-cost provider that runs the model well.
2. If the cheapest route lacks the model, has no key, fails auth, or errors, fall back to the next provider.
3. Never hide a provider swap. Say which route ran and why.

Usage

python3 execution/generate_media.py list                    # routes + prices + key status
python3 execution/generate_media.py --prompt "..." --dry-run  # quote, spends nothing

# default: prepaid 1min.ai, no real money
python3 execution/generate_media.py --prompt "dark charcoal banner, lime glow" --project deece

python3 execution/generate_media.py \
  --prompt "wide shot of a lime gradient key visual, dark background" \
  --model nano-banana --project deece --aspect 16:9 \
  --ref generations/refs/deece-logo.png

python3 execution/generate_media.py --prompt "slow push in" \
  --model veo --duration 8 --yes        # --yes required for anything paid

--dry-run calls nothing and writes nothing. Use it to quote.

Output

• Every file lands flat in generations/. No subfolders, ever.
• Reference images live in generations/refs/ and are the only thing in there.
• Naming: {project}_{prompt-slug}_{timestamp}.{ext} — the script does this.
• The gallery at generations/gallery.html rebuilds after every save. Open it in a browser; it needs no server.
• Generated media is gitignored. The sidecar .json logs and refs/ do sync, so the prompt library survives across the Mac and the PC.

Rules

1. Quote the cost and wait for Drew's explicit go before any paid video run. Quoting is not approval. One approval = one run. The script enforces this in two currencies: it refuses anything over $0.05 of real money, or over 50,000 prepaid credits, without --yes.
2. Draft cheap, finish pretty — and prefer prepaid over paid. Iterate on flux-schnell (prepaid, no real money). Only rerun the same prompt on a quality model once Drew picks a favourite. Reach for a real-money route only when the job needs something prepaid can't do: references, or text in image. Same for video: prove the motion on veo-fast before paying for veo.
3. Never describe a logo or a face in text. Pass the real file with --ref. A described logo comes back wrong every time — wrong shape, wrong colours, invented detail. If the reference is missing, stop and ask Drew for it. (See also the memory note on AI likeness from reference photos: pin the person count, never describe the face.)
4. One generation at a time. Run them sequentially to avoid rate limits.
5. Never invent a model id. A 404 means the provider shipped a new version: open their model page, copy the id fresh, update MODELS in the script and the recipe file. That is the only maintenance this system needs.
6. After every save the script writes the sidecar log. Don't skip it and don't write it by hand.
7. Bake off the same prompt before committing a model to a job. For a hero shot or any clip with on-screen text, run one identical prompt on two or three video models (same seed idea, same ref) and compare lettering, hands and realism before the full run. Price is not quality: in one public benchmark a Kling 3.0 clip at about $0.60 matched a Seedance 2.5 clip at about $3.23 on selfie-style UGC, while Seedance won on legible text. Say each run's cost in the report so the comparison is visible. (source: Nate Herk, "This ONE GPT-6 Astra Skill Replaces Your Higgsfield Subscription", ingested 2026-10-01)
8. Per-use API vs flat subscription is a break-even call. Below roughly 16 to 50 clips a month (varies by plan) pay-per-use wins, above it a subscription does, and anything programmatic or bulk favors the API. Higgsfield's pay-per-use API is wired in as of 2026-10-02 (kling verified live, seedance-2.5 verified live 2026-10-08; recipe models/higgsfield.md); compute the break-even against our real monthly clip count before leaning on it, and remember Drew does not spend money on trials. (source: same, ingested 2026-10-01)

Adding a model

1. Add a row to MODELS in execution/generate_media.py (alias, provider, model_id, kind, cost, verified=False).
2. Copy models/_template.md, fill it from the provider's docs.
3. Add a row to the table above.
4. First successful real run: flip verified=True and record the confirmed response shape in the recipe.

Related

• YouTube thumbnails: don't use /generate directly. Read directives/thumbnail_forge.md and use execution/thumbnail_forge.py, which wraps this script in a generate -> critique -> regenerate loop and refuses to hand back an image the critic has not passed. execution/thumbnail_gen.py and execution/wealth_strategist_thumbnail.py remain the right tools for a fixed template. /generate is for everything else.
• Any generated image that ends up on a page or in an email still has to pass /test-improve before it ships.