Article covers: when they are drawn and how they are checked

If your site shows a cover image with each article, the content engine gives every article one. AutoPSEO's writer draws the cover while it writes, in the style of your existing covers, and a vision model throws away any picture with a logo or readable text. Your own AI can bring its own image instead. Covers are 1536 by 1024 pixel JPEGs, delivered with the article, and they use neither your article allowance nor your research credits.

When an article gets a cover

AutoPSEO draws covers only for a site that shows them:

• Your site renders the articles through the webhook, and has said it shows covers: its ping answer carries covers: true, or it answered a delivery with cover: "wanted" at least once. From then on, AutoPSEO's writer draws the cover right after it saves each draft, so the article's first delivery already carries it.
• An article arrived without one. Answer that delivery with cover: "wanted": the five-minute schedule draws one, oldest request first, and sends it to you in an article.updated.
• AutoPSEO hosts the articles and the template it learned from your site shows covers (hosting).

A site that shows no covers gets none. The five-minute schedule gives up on an article after three failed attempts. A live article keeps its cover until someone replaces it with content_set_cover; a revision waiting for review never changes it.

How a cover is drawn

1. The scene. A text model on Workers AI (Llama 3.3) describes one scene of at most 45 words for the article: generic people, the place, what they do, the objects around them. It names no company, brand, product, app, website, school, person or place, and puts nothing readable in the picture. A scene that still names your brand is refused. The image model never sees the article itself.
2. The picture. FLUX.2 [klein] draws the scene "in exactly the style, lighting, palette and framing" of up to two of your own share images or covers, used as references. Without references it draws a calm, natural-light editorial photograph. The prompt asks for no text, letters, numbers, logos or watermarks, abstract shapes on any screen, and plain unbranded laptops, phones and monitors.
3. The check. A vision model (Llama 4 Scout) looks at the result. A picture with a logo or readable text counts as a failed attempt and is drawn again on a later run.

The writer's cover step is never retried on its own. If it fails, the article carries on without holding anything up.

The file your site receives

Deliveries carry it as article.cover with url, sha256, contentType, width and height. Fetch it once and keep your own copy; your site owns the image like the rest of the article. On AutoPSEO hosting, covers are served under the content type's path at /_cover/<slug>.jpg, in the sizes and formats your template uses.

Bring your own image

A writer working over MCP can supply any image it likes, made however it likes (your own AI):

• With the draft: content_submit's cover, either url (an https address on the public web, fetched once) or data (base64, up to 2.5 MB).
• As a file: a PUT of the image bytes to the coverUploadUrl that content_submit returns when your site wants a cover.
• Later: content_set_cover with an image, or with none to get an upload address. This is also how a live article's cover is replaced.

AutoPSEO checks your image only as a picture: PNG, JPEG or WebP, between 10 KB and 12 MB, at least 800 pixels wide and 400 high, and landscape (no taller than wide, no wider than 2.5 to 1). It is centre-cropped to 3:2 and saved as a 1536 by 1024 JPEG. The vision check does not run on your own image, and it is kept even when AutoPSEO does not know whether your site shows covers: your site decides.

Draw from AutoPSEO's prompt

A writer with an image tool of its own can still use AutoPSEO's prompt and checks:

1. Describe a scene of 20 to 600 characters and send it with content_cover. AutoPSEO checks it like its own scenes and returns the image prompt, up to two of your site's covers as references and a signed upload address.
2. Draw the picture from that prompt with any image tool, and PUT the bytes to the address. It works for one hour and takes up to five pictures.
3. AutoPSEO resizes the picture to 1536 by 1024, runs the vision check and keeps it. A live article is sent to your site again with its new cover.

The unattended runner described in your own AI uses this flow with its make_cover tool, drawing with the Codex CLI or the Grok CLI on your own Mac.