| name | manga-image-prompt |
| description | Generate OpenAI/GPT Image 2 compliant manga page prompts from a page script, project DESIGN.md, and character references. Use for direct final page render prompts, repair prompts, reference indexing, English/direct text policy, two-round prompt review, and quota-aware render preparation. Storyboard boards are not required. |
Manga Image Prompt
This skill writes the prompt that may spend image quota. Be conservative.
Required Inputs
Read design/DESIGN.md and the relevant page from script/page-script.md. Use character reference images when available and name each one by index. Do not require a board PNG or board JSON.
Default final-page references are character-only. If the user explicitly asks to pass scene or prop references, name each one by index and state exactly what it controls. Otherwise, let the model generate environments, props, camera angles, and panel layout from the page script and DESIGN.md.
Use templates/page-render-prompt.md for final pages and templates/prompt-review.md for the two-round review.
Prompt Order
Follow a stable OpenAI-friendly order:
Intended use and output mode
Reference images by index and authority
Project DESIGN style lock
Page story function
Continuity and object locks
Suggested page rhythm and key visual moments
Dialogue, captions, SFX
Text policy
Change / preserve contract
Technical flags to use
This mirrors the practical prompting fundamentals: clear structure and goal, concrete medium and materials, explicit composition, pose and object interaction, clear preserve/change constraints, exact text in quotes, and reference-image authority.
Monochrome Reference Rule
For black-and-white manga, every prompt using colored references must include this rule in plain language: ignore all hue, saturation, palette, colored light, and gradient information in reference images. Preserve only identity, silhouette, construction, costume layering, prop geometry, spatial layout, material wear, and pose logic. Final output is monochrome ink on off-white paper.
Direct English Text Rule
When the user asks for English dialogue directly in the image, put every literal text item in quotes and state that no other text may appear. Keep text count low. For talismans, slips, ledgers, shop papers, labels, and signs that are not explicitly quoted, require blank paper, ruled lines, abstract seals, or non-language ink strokes. This prevents pseudo-Chinese or random letters from leaking into props.
If exact lettering is more important than speed, write a no-text or blank-balloon render prompt and route final lettering outside image generation.
Two-Round Review
Round 1 reviews story and manga function: page goal, suggested rhythm, reader eye path, action clarity, dialogue function, object continuity, text economy, and page-turn value.
Round 2 reviews OpenAI and wrapper compliance: skimmable prompt structure, reference indexing, exact text handling, preserve/change instructions, negative constraints, monochrome style lock, and flags such as 1088x1536, png, and quality high.
Do not render before both rounds pass. If either round fails, revise the prompt instead of spending quota.
