History Comics Produce

History Comics — Produce (Content + Topview)

This skill is the orchestrator. It does not replace the other two skills.

Read both sibling skills before executing. Never invent Topview REST calls or run topview-generate/scripts/*.py. Use host MCP tools from topview-mcp (topview_*, ta_*) exactly as topview-generate specifies.

Do not use this skill for Amazon / Shopee / TikTok Shop / YouTube research.

Target layout

Every produced comic must end as:

content/<country>/<period>/<slug>/
├── meta.json
├── story.md
├── story.en.md
├── prompt.md
├── images/
│   ├── 00-character-sheet.png
│   ├── 01.png
│   ├── 02.png
│   └── manifest.json
└── comic.png                 # website cover = panel 01

comic.png stays at the comic root because the current site reads /content/.../comic.png and hasImage is true when any image file exists in the comic directory.

When to run which phase

If prompt.md is missing or still a legacy one-paragraph stub, run Phase A first. Do not invent image prompts ad hoc.

Phase A — Author the entry

Follow history-comics-content in full. Stop after its validation checklist passes. Then continue to Phase B only when the user asked for images or a full pipeline.

Phase B — Generate images with Topview

Follow topview-generate Common Agent Workflow for every MCP action: connect → auth if needed → pick/create board → live generation config → submit → poll the same taskId → return URLs.

B0 — Preconditions

1. Confirm topview-mcp is connected. If discovery fails or tools are missing, invoke host mcp_auth. If it returns a URL, send that exact URL and wait for the user to confirm sign-in. Do not switch to another image API.
2. Read prompt.md. Extract, in order:
   • every fenced English Prompt (Copy below) block (these are the only prompts that may be submitted);
   • the character English descriptors from 分镜规划;
   • the count-sensitive prop ledger.
3. Call topview_get_generation_config with type=image for both text_to_image and image_edit. Prefer GPT Image 2 when present (best bilingual text). Use models[].submitModel exactly.
4. Call topview_get_credit before the first chargeable submit. Present one plan: panel count, model, 16:9, resolution (2K if supported), estimated credits, and whether later panels may auto-proceed. “Just do it” / “直接出图” counts as auto-proceed for the rest of this comic.
5. Reuse one board for the comic. Prefer an existing board named history-comics / <中文标题>; otherwise topview_create_board with that name.

B1 — Character sheet (required lock)

Before any story panel, generate one character-sheet image.

• taskType: text_to_image
• Prompt: same global style + period palette as prompt.md; all named characters standing in a row, full-body, neutral stance, no speech balloons, no extra extras, each with the exact bilingual name label from the Series Bible (晏子 / Yan Zi). No story action.
• aspectRatio: 16:9 (or the closest live-config value)
• generateCount: 1

Poll to success, download the image to images/00-character-sheet.png, then upload it with the Topview upload flow (ta_upload_credential → PUT bytes → ta_upload_check_file). Keep the resulting fileId as characterSheetFileId for every later panel.

Do not skip the sheet. Text-only character descriptions are not enough for cross-panel consistency.

B2 — Panels in order

For panel N of 1..K:

1. Take the Nth English Prompt verbatim. Do not shorten it.
2. Submit image_edit (not plain text-to-image) with:
   • prompt = that English Prompt
   • inputImageFileIds = [characterSheetFileId] plus the previous accepted panel fileId when N > 1
   • aspectRatio: 16:9
   • generateCount: 1
   • same boardId and model as the sheet
3. Poll topview_query_task with the same taskType and taskId until success or fail. Timeouts are not failures; keep polling.
4. QA the result against the panel's Chinese description and prop ledger (character colors, peach/sword counts, no gore, bilingual text present). One image_edit revision is allowed per panel if a ledger item is wrong. After a second failure, stop that panel, keep the last URL, and continue.
5. Download the accepted image to images/0N.png (01.png, 02.png, …). Upload it if a later panel will reference it and you do not already have a reusable fileId.
6. Independent panels must still run in order. Do not parallelize story panels; later frames depend on earlier fileIds.

If live config has no image_edit, fall back to text_to_image with the same verbatim prompt, and say so in the handoff. Still generate the character sheet first so a later edit pass can lock likeness.

B3 — Save locally

Write images/manifest.json:

{
  "boardId": "",
  "model": "",
  "generatedAt": "YYYY-MM-DD",
  "characterSheet": {
    "local": "images/00-character-sheet.png",
    "fileId": "",
    "taskId": "",
    "url": ""
  },
  "panels": [
    {
      "index": 1,
      "local": "images/01.png",
      "fileId": "",
      "taskId": "",
      "url": ""
    }
  ]
}

Copy images/01.png to comic.png at the comic root (website cover). Do not overwrite an existing comic.png unless this run produced a new panel 01.

After files are on disk, tell the user to run cd web && npm run build:content if they want the site index to pick up hasImage.

B4 — Handoff text

Keep the user-facing summary short. For each panel list the local path and the download URL. Include the board link only when both boardId and boardTaskId are known:

https://www.topview.ai/board/{boardId}?boardResultId={boardTaskId}

Hard rules

• Do not submit a shortened or rewritten English Prompt.
• Do not generate gore, weapon-to-body contact, corpses, or scary close-ups. If Topview returns safety code 6001, revise only the unsafe clause and retry once.
• Do not resubmit a still-running task.
• Do not route Topview creative tools through get_tool_schema / call_tool.
• Do not modify the official topview-generate skill. If its MCP sequence changes, follow that skill, not this file's examples.

Current site gap

The website still shows a single comic.png. Multi-panel files in images/ are the source of truth for production. Do not change web/ from this skill unless the user explicitly asks to render a panel gallery.