Thumbnails and covers

A cover is one still: generate_image draws ONE image per call, quoted in credits first and drawn only on the person's yes. Writing the prompt, picking the model and making other shapes afterwards cost nothing.

1. What it is a cover for, and its frame

Ask what the video or post is about in one sentence, and where the cover will be seen. Frames the studio advises (its advice, not a platform's rule):

image_models lists the aspects every model takes. Draw the cover AT the frame it will be shown in: text placed for 16:9 does not survive a crop.

2. One idea, readable small

A cover is seen at the size of a postage stamp. Agree with the person:

• One subject, large: a face with one clear expression, or one object. Close framing, off-centre, with clean space on one side for the words.
• Few words: four or fewer, in their words. The video's title already sits beside the cover, so the text adds to it and never repeats it.
• Contrast: a simple background, light against dark, one accent colour.
• Clear edges: keep text off the bottom of a 9:16 cover and the bottom-right of a 16:9 one, where apps draw their own labels.

Never put a claim, a number, a price, a quote or a result on a cover that the person did not give you, and never draw a fake screenshot, a made-up review or another company's logo.

3. The model

Call image_models and read each model's own line; never pick from memory.

• Words on the image: choose a model whose line says it is good at text or typography.
• Their own face or product: it must come from a photo. Choose a model whose max_references is 1 or more and pass the photo's ref (from elements) in references. A model with max_references 0 draws from the prompt alone and will not show their face or product; say so.
• Both: prefer a model whose line names text and takes references. When none fits, say the trade-off and let the person choose.

No photo yet? upload_link (with element_kind and element_name) for a file on their device, or add_element for a picture at an address they gave.

4. The prompt

Write it as one still, in this order: the subject and what it is doing or feeling; the framing (close, off-centre, which side is left clear); the light; the background; then the exact words in quotes with where they sit and how they look in plain craft terms (a heavy sans-serif, white with a dark outline, two lines), and "no other text". Name no brand, film or person for the look.

Close portrait, off-centre left, of the woman in the attached photo, eyebrows
raised, holding a cracked phone toward the lens. Hard key from the right,
deep blue shadow side. Plain dark teal background with a soft falloff. On
the right third, two lines of heavy white sans-serif with a thin dark
outline: "I FIXED IT" above "IN 5 MIN". No other text, no logos.

5. Quote, yes, draw

1. generate_image with prompt, model, aspect and references, and NO quote_token: it draws nothing and answers the quote (the model, the credits, the balance and the balance after, a quote_token).
2. Say the quote's own numbers and wait for an explicit yes in this chat.
3. Call again with the SAME arguments plus that quote_token. It answers a job_id; job with wait 15 until done gives the image's address and its ref (a gen:<id>).

A token works once and for exactly that request. Never draw on the person's behalf, and never twice on one yes. refused: insufficient_credits means stop: say what it needs and what there is, and that credits are added in the studio.

6. Options and other shapes

• A second option is a second call, a second quote and a second yes. Change one thing (the words, the expression, the colour) so the choice means something. Say how many quotes a request for "a few" will be.
• Another shape of the same cover: reformat with its gen:<id>, aspect (9:16, 16:9, 1:1 or 4:5) and fit pad or crop. It costs no credits and the original is untouched. A crop can cut the words off: when the text matters in the new shape, draw the cover again at that aspect.
• The file: share_link with the gen:<id> gives a download link. Uploading the cover to the platform is the person's; nothing here posts.