| name | manga-illustrator |
| description | Create a reusable visual package for manga video frames by default, or pages/panels only when explicitly requested, with stable character identity and prompt discipline. |
Manga Illustrator
Use this skill to create the reusable visual package for the manga video studio team.
Expected Inputs
series-bible.md
character-bible.md
character-registry.md
series-state.md
chapter-registry.md
chapters/<chapter-id>/chapter-plan.md
chapters/<chapter-id>/storyboard.md
chapters/<chapter-id>/continuity-ledger.md
storyboard.md is the operative source of truth for what must be rendered. chapter-plan.md is only supporting scope context once the storyboard exists.
The storyboard must also tell you whether this run is video-frame, page-composed, panel-first, or key-asset preview.
Produced Artifacts
chapters/<chapter-id>/visual-style-guide.md
chapters/<chapter-id>/prompt-pack.md
chapters/<chapter-id>/image-generation-log.md
shared-assets/character-sheets/character-*.png
chapters/<chapter-id>/frames/frame001.png ..., chapters/<chapter-id>/pages/page01.png ..., or chapters/<chapter-id>/panels/panel001.png ...
chapters/<chapter-id>/visual-production-package.md
Use visual-style-guide-template.md to structure the chapter visual guide.
Use prompt-pack-template.md to keep durable prompt blocks consistent across chapters.
Use visual-production-package-template.md as the delivery skeleton.
Also use image-generation-log-template.md for durable image-tool bookkeeping.
Required Shared Reads
- Start by reading production-principles.md.
- Use it as the shared reference for visual mode, character locking, and reproducibility.
Workflow
Step 1 - Lock the visual language
Write chapters/<chapter-id>/visual-style-guide.md before bulk generation.
Start from the series-level visual identity in series-bible.md. The chapter guide should inherit that canon first, including the locked series video aspect ratio, then add chapter-specific execution notes. Chapter-local style notes must not redefine the video canvas.
Record:
- inherited series look promise
- chosen visual mode
- color strategy for this chapter
- linework and contrast rules
- screentone or shading rules
- realism versus caricature guidance
- expression and exaggeration guidance
- camera grammar
- background rendering rules
- typography or SFX style notes
- inherited locked series video aspect ratio and target resolution when video delivery is in scope
- negative prompt block
- consistency notes for recurring characters and locations
If the user has not specified a style, default to a readable manga look rather than a painterly poster style.
Do not casually override a locked series identity. If the series is full color, comedic, airy, grotesque, or hyper-dramatic, the chapter guide should preserve that instead of collapsing into generic grayscale tension manga. Never override the locked series video aspect ratio inside a normal chapter.
Step 2 - Generate character reference sheets first
- Use
generate_image as the default image-generation route.
- Use
edit_image as the default local-fix route when an otherwise strong image needs targeted repair.
- For prompt-only
generate_image routes, omit generation_config and do not specify a model identifier. Use the configured default image route. The final prompt must be self-contained.
- Issue image-tool calls strictly sequentially. Call exactly one
generate_image or edit_image, wait for the result, inspect and log it, then immediately run sleep 60 before any further image-tool call.
- The same cooldown applies after successful, rejected, timed-out, and failed image-tool calls. Do not retry before running
sleep 60.
- Do not launch multiple
generate_image or edit_image calls concurrently, in the background, or as a parallel batch. This applies to reference sheets, video frames, pages, panels, retries, and local fixes.
- Create one reusable reference sheet per recurring character before rendering the main video frames. Render pages or panels only when the declared non-default contract explicitly requires them.
- Use the lock language from
character-bible.md.
- If the chapter introduces a new recurring character, create or extend that character's reusable sheet package immediately after the design is approved.
- Feed approved character sheets and other locked references back into later
generate_image calls through input_images when the scene needs those references.
- For scenes with multiple recurring characters or important recurring locations, include the relevant approved image paths in
input_images instead of hoping the prompt alone will preserve identity.
- If the active runtime expects multiple
input_images refs as a comma-separated value, build that argument in the runtime's required comma-separated form.
- Store approved reusable sheets in
shared-assets/character-sheets/ so later chapters can start from them.
- Do not move into main frame generation if the character identity is still drifting. The same gate applies to page/panel generation when explicitly requested.
Step 3 - Write chapters/<chapter-id>/prompt-pack.md
Create a durable prompt package that includes:
- one canonical character block per recurring character
- one location block per recurring setting
- one video-frame prompt per storyboard unit for video delivery, or page/panel prompts only when explicitly requested
- lettering specs for any video frame that must contain readable on-image text, or for page/panel text only when those assets were explicitly requested
- negative prompts for drift, extra limbs, wrong props, wrong clothing, wrong age presentation, and unreadable composition
Every final prompt sent to generate_image must be self-contained for the selected render-unit contract. It should include:
- asset id or render-unit id
- render-unit contract and art state, such as
video-frame, page-composed clean-art base, panel-first clean-art, or lettered panel
- aspect ratio and orientation in prompt text, such as
9:16 vertical, 16:9 horizontal, or vertical 2:3 portrait
- scene and visible action
- character lock details needed for identity
- style and color mode
- text strategy, such as exact readable text, clean empty areas for later text, or
do not render readable text
- negative constraints such as no watermark, no signature, no unwanted text, no wrong character type, and no contract-breaking layout
When a video frame must carry readable text, the prompt pack must specify the fields below. Apply the same fields to page/panel text only when an explicit non-default contract requires those assets.
- exact text to render
- whether it is a speech bubble, caption box, SFX, signage, or chapter-end mark
- language or script
- approximate placement on the frame, or on the page/panel only when those assets were explicitly requested
- reading order when more than one text element appears
- lettering style notes when that affects readability
For a video-frame contract, every render-unit prompt must specify:
- inherited locked series aspect ratio and orientation, such as
9:16 vertical or 16:9 horizontal
- target canvas or dimensions in prompt text when known, such as
1080x1920 vertical
single full-bleed motion-comic frame
- one visible beat or one coherent full-screen composition
- subject-safe and subtitle-safe composition notes
- explicit negative prompt language:
no manga page, no panel grid, no page border, no white paper margin, no collage, no contact sheet, no reference sheet, no watermark
The Exact final prompt entry in prompt-pack.md is the actual constructed prompt to send to generate_image.
For video-frame delivery, that final prompt is invalid unless the prompt text itself contains the locked series aspect ratio and orientation, such as single full-bleed 9:16 vertical motion-comic manga frame.
For page-composed delivery, that final prompt is invalid unless the prompt text itself contains the approved page aspect ratio and orientation, such as vertical 2:3 portrait full-color manga page clean-art base.
For panel-first delivery, that final prompt is invalid unless the prompt text itself contains the approved panel shape or framing intent.
Do not rely on metadata fields or generation_config to carry the aspect ratio, orientation, or target dimensions.
Do not ask the image model for "two panels", "three panels", "manga page layout", "clean manga borders", or similar page-composition language when the contract is video-frame.
If the storyboard beat contains multiple actions, split it into multiple video frames or stage a single coherent full-screen shot. Do not solve it by arranging several small images in rows on a paper-like background.
Do not rely on bracket notation such as [laughs] as if it were special image syntax.
For image generation, describe the visible result directly.
If a character should look like they are laughing, describe the visible facial and body cues.
If the image should contain text, say exactly what text should appear and where.
Examples:
Good video-frame prompt fragment:
single full-bleed 9:16 vertical motion-comic manga frame, Ada stepping from the train into amber station light, visibly focused young scholar, red notebook visible, immersive cinematic composition, subtitle-safe lower band with low visual clutter, no manga page, no panel grid, no page border, no white margin, no collage, no watermark
Good page-composed clean-art prompt fragment, only when the contract is `page-composed`:
vertical 2:3 portrait full-color manga page clean-art base for ch002-p01, one lush establishing splash page, midnight-blue and brass fantasy train gliding into a station under high arches, deep burgundy carpets hanging like banners, clerks in felt slippers crossing silently, brass lamps glowing through dust, compact badger scholar in midnight-blue coat with burgundy scarf, round spectacles, red notebook, standing alert in the train doorway, slim white cockatoo with yellow crest and citron waistcoat peering out excitedly, premium full-color animal fantasy manga, crisp ink lines, cinematic station architecture, leave clean empty areas for a station sign, banner, one speech bubble, and tiny SFX, do not render readable text, no grayscale, no human characters, no watermark, no signature
Good panel clean-art prompt fragment, only when the contract is `panel-first`:
black-and-white manga panel, Morena leaning forward over the card table, cold smile, Borksen rigid across from her, tense silence, dense screentone shadows, no speech bubbles, no captions, no watermark
Good panel lettered prompt fragment, only when the contract is `panel-first` or `page-composed`:
black-and-white manga panel, Kurapika at the doorway checking the list, Borksen visible at the end of the corridor, one clear speech bubble near Kurapika containing the exact Chinese text "下一位。", one second speech bubble near Borksen containing the exact Chinese text "……我是来参加讲习的。", readable handwritten manga lettering, right-to-left bubble flow, keep text large and legible
Step 4 - Log every material image-tool call in chapters/<chapter-id>/image-generation-log.md
For every material generate_image or edit_image call that affects an approved or candidate asset, record:
- asset id
- tool or image route used
- model identifier returned by the tool, if available
- exact final prompt or edit instruction sent to the image tool
- applicable aspect ratio and orientation phrase present in the final prompt
- visible text specification or explicit clean-art instruction
- source image refs
generation_config status, which should be omitted for prompt-only generate_image routes
- actual output dimensions and measured aspect ratio
- output path
- approval status
- sequential-call position and whether
sleep 60 was completed before the next image-tool call
- reason for the iteration
Future chapters should be able to reuse or refine these exact records instead of guessing what happened.
Step 5 - Generate the main visual assets
- Render the required video frames in storyboard order. Render pages or panels only when the declared non-default contract explicitly requires them.
- Generate assets one at a time in storyboard order. After each
generate_image or edit_image result, timeout, or failure, immediately run sleep 60 before requesting the next asset, retry, or correction.
- Follow the declared chapter render-unit contract exactly.
- If the contract is
key-asset preview, treat the package as incomplete chapter coverage unless the user explicitly asked for preview-only output.
- If the contract is
video-frame, every approved asset must be one full-bleed video still at the inherited locked series aspect ratio, with no page border, panel gutter, paper margin, collage layout, reference-sheet layout, or generated watermark.
- If the contract is
page-composed, do not stop at scene key art or character locks.
- If the contract is
panel-first, do not stop until every renderable storyboard panel has a final approved asset.
- Keep filenames stable and production-friendly.
- Reuse reference sheets, recurring location blocks, and prior approved images to reduce drift.
- Prefer
generate_image with deliberate input_images over blind prompt-only regeneration when identity consistency matters.
- If the result needs small local fixes, use
edit_image instead of regenerating the entire package blindly.
- Do not invent staging, dialogue, or emotional beats from
chapter-plan.md when storyboard.md is already explicit.
- For
video-frame delivery, inspect each generated frame before approval. Reject and regenerate frames that read as a manga page on paper, a multi-panel sheet, or an image pasted onto a blank background.
- Make the lettering state explicit for the delivered assets:
clean-art
partially-lettered
fully-lettered
Step 6 - Write chapters/<chapter-id>/visual-production-package.md
Record:
- chapter render-unit contract
- chapter visual delivery status
- package lettering state
- subtitle overlay expectation for downstream video use
- visual style summary
- asset manifest
- approved storyboard-to-asset mapping
- reusable shared-asset updates
- new or updated character-sheet assets
- character sheet paths
- video-frame paths for video delivery, plus page/panel paths only when explicitly requested
- inherited locked series aspect ratio and frame-conformance QA for
video-frame delivery
- prompt-pack summary, including confirmation that final prompts contain the applicable aspect ratio and orientation phrase
- image-generation-log path
- QA findings
- unresolved visual risks
The approved storyboard-to-asset mapping should make it unambiguous which final image path corresponds to each storyboard video frame, page, panel, or other renderable unit.
It should also make it unambiguous whether the mapped asset is clean art, partially lettered, or fully lettered, and whether the video stage should burn subtitles over it.
If any storyboard render unit is not covered, keep the package marked incomplete and list the missing units directly.
Then hand off the full cumulative package to voice_video_producer.
Blocking Rules
- If the canon does not lock the character well enough to render consistently, route the issue upstream.
- If a key visual moment cannot be rendered clearly from the storyboard, route that issue back to
storyboard_director.
- If the output drifts repeatedly in ways that break identity, stop and tighten the style guide and prompt pack before continuing.
- If the delivery target is a full chapter and the package only contains preview art or partial coverage, do not represent it as finished chapter imagery.
Illustrator Standards
- Consistency is more important than novelty after the visual language is locked.
- Prefer clear storytelling images over highly detailed but unreadable compositions.
- Leave behind a prompt package that can actually support later chapters.