Ad creative

Turn a product into ad creatives that a buyer could actually run: a stated angle, a hook written to be read in under a second, and a rendered image sized for the placement.

Requirements

The inReels MCP server must be connected. Check with get_account — if it fails, the user needs to add the connector:

claude mcp add --transport http --scope user inreels https://www.inreels.ai/api/mcp

Then set an API key from https://www.inreels.ai/dashboard/settings/api-keys.

Method

1. Get the angle before the image

Do not generate anything yet. Ads fail on the angle far more often than on the render, and an image made before the angle is settled is a wasted credit.

From the product (URL, description, or images the user gave you) establish:

• Who it is for, specifically. "Busy people" is not a who.
• The friction they feel today, in their words, not the brand's.
• The angle — the single claim this creative makes. One per creative.
• The hook — 3 to 7 words, readable at a glance in a feed.

If the user gave you a landing page, read it first. The page's own headline is usually the weakest available angle, because it was written to satisfy everyone. Look at what the page's objection-handling and FAQ reveal about the real hesitation, and aim there.

State the angle and hook back to the user before spending anything. One round of correction here is worth three regenerations.

2. Price it

quote { "media": "image", "image": { "prompt": "<your prompt>", "aspect_ratio": "9:16" } }

The number returned is exactly what the generation charges. Report it if the user has not already approved a spend.

3. Choose the model deliberately

Call list_models and pick on what the creative needs:

• `nano-banana` — the default. Good general renderer, accepts up to 3

reference images, so this is the one to use when the real product must appear.

• `nano-banana-pro` — when the creative carries text or needs detail to hold

up at full size.

• `flux-schnell` — cheapest. Use it to test several angles quickly before

committing to a final render.

• `flux-kontext-pro` — editing. Requires exactly one reference image, and

edits it rather than inventing a new scene.

If the user supplied product photos, pass them as ref_urls so the product in the ad is the real product. An invented stand-in is the fastest way to make a creative unusable.

4. Render

generate_image {
  "prompt": "<scene, framing, lighting, where the product sits, where text goes>",
  "aspect_ratio": "9:16",
  "model": "nano-banana",
  "ref_urls": ["https://…/product.jpg"]
}

This returns a run_id immediately. Then:

get_run_status { "run_id": "…", "wait": true }

wait: true blocks up to 45 seconds and usually returns a finished image in one call. Never give the user a URL before the run reports completed.

5. Vary one thing at a time

When the user wants options, change exactly one variable per creative — the angle, or the framing, or the product placement — and say which. Four images that differ in everything teach nothing about what worked.

Prompting for ads specifically

What actually moves ad performance, in rough order:

• Leave room for text. Decide where the hook will sit and keep that area

uncluttered. State it in the prompt: "lower third clear for text overlay".

• Native over polished. Feed ads that look like studio work read as ads and

get scrolled. Phone-camera framing, ordinary light, and a real surface beat a seamless gradient.

• Product legible at thumbnail size. If it is unrecognisable at 100px wide,

the creative does not work, however good it looks full size.

• One subject. Two competing focal points halves both.

Spending rules

• quote before generate_image whenever the user has not pre-approved.
• Failed runs are refunded in full, so a failure is not a loss — but a

successful render of the wrong angle is.

• Do not loop. If two renders in a row miss, the prompt is wrong, not unlucky;

go back to the angle and say so.