blue-poster

Blue Poster

Generate each final poster from the original source image in one image-generation call. Default to a complete full-design poster. Never create a base image and then add effects in a second content-generation stage.

Required tool policy

• Detect the host environment before generation and use only an image-generation capability built into that host.
• In Codex, use only the Codex host built-in image_gen.imagegen tool. Never invoke, install, or fall back to a third-party image skill, Ark, Doubao, Gemini, a local model, CLI, or API workflow.
• Outside Codex, use the current model or platform's own built-in image-generation capability when it is available, such as GPT's native image generation in ChatGPT or Doubao's native image generation in Doubao. Do not call another provider's API, CLI, local model, or third-party image skill as a fallback.
• If the current host has no built-in image-generation capability, or that capability fails, stop and report the blocking reason.
• Treat style reference images as browse-only. Never pass them to image generation.

MPO source normalization

• .mpo is an accepted source container. MPO is normalized before generation; it is not a new poster output mode.
• Inspect frames with python3 scripts/extract_mpo.py <source.mpo> --list.
• Unless the user chooses another frame, extract frame 0, the primary JPEG, and use that normalized JPEG as the only image input to the selected host-native image generator.
• Do not merge stereo frames or invent depth semantics. If the user asks for left/right or multi-view treatment, clarify the selected frame before generation.
• This local extraction step only unwraps the source container. It does not generate, alter, crop or stylize image content.

Generate the default full poster

1. Require one source image. If it is a local file, inspect it with view_image before generation.
2. If the user names S01–S11, honor it. Otherwise read style-selection.md, analyze the source semantics, and select exactly one style.
3. Read prompt-full-design.md and the selected block in style-programs.md.
4. Replace {{EFFECT_PROGRAM}} with exactly one complete style block. Do not leave placeholders or merge multiple styles.
5. Call the selected host-native image generator once with the original source as the only image input. In Codex, this means image_gen.imagegen. State that the source is a semantic and structural source, not a style reference.
6. Save the selected image under the caller's current workspace at outputs/blue-poster/<source>-full-Sxx-vN.png, unless the user supplied another destination. Never overwrite an existing file.
7. Save the exact final prompt beside it as <source>-full-Sxx-vN.prompt.md.
8. Run python3 scripts/validate_output.py <image> --mode full. If it fails, retry once from the original source with the same style and a stronger native 3:4 instruction. Never stretch or crop to hide a ratio failure. If the retry fails, report it as unaccepted.
9. Perform only a lightweight check on the returned image: the subject remains recognizable, the whole page is designed, no photographic window remains, and the full-mode carrier contract is visible—3–6 active regions, no more than two clusters, no isolated small rectangle, a continuous subject corridor and at least 40% quiet paper. Do not start a browser or create extra screenshots.
10. Report the selected style ID and name, one concrete selection reason, image path, prompt path, validation result, and catalog path.