cmm-visual-production

name	cmm-visual-production
description	Use this CMM skill to plan visual production for content packages, including WeChat covers, Xiaohongshu/Rednote card sets, article illustrations, social cards, thumbnails, layout specs, and image prompts. It directly creates visual_plan and visual_package contracts with CMM asset specs, prompts, and quality gates, while treating legacy visual skills only as optional approved execution backends. This new CMM layer does not modify legacy visual skills and does not silently generate images.
version	0.2.0

CMM Visual Production

cmm-visual-production is the visual routing layer in the CMM content factory.

It consumes a content_package, creates a visual_plan, and produces visual asset specs, prompts, layout guidance, and quality checks. After user confirmation, an approved backend can render or generate the assets. Its final artifact is a visual_package.

Boundary

Use this skill for:

• Planning visual assets for platform drafts.
• Deciding whether the content needs a cover, article illustrations, social cards, Xiaohongshu images, thumbnails, or prompts.
• Producing asset specs, image prompts, layout notes, and quality gates.
• Selecting an optional execution backend after confirmation.
• Producing a structured visual_package for publishing workflows.

Do not use this skill to:

• Write the content draft.
• Render or generate images without confirmation.
• Patch or modify generated bitmaps.
• Publish or upload assets to external platforms.

Generation Boundary

This skill plans visuals and produces a visual_plan / visual_package (covers, body illustrations, social cards, Xiaohongshu card sets). Actual image generation is performed by cmm-image-creator after confirmation — and that confirmation must include the locked 文案三件套(标题/正文/图上要印的逐字文案)； 生图前置确认门见 cmm-image-creator/SKILL.md →「生图前置确认门」(硬规则)。All visual methods (illustration framework, density tables, platform specs, card layouts) are internalized in this skill's references/ and templates/ — no external skill is required.

For WeChat covers, readable baked-in title text is mandatory. Do not plan or accept a cover whose policy is no baked text, title handled by metadata, 无字底图, 无文图, or any equivalent no-text cover. Image content and text must be generated by the approved image-generation backend. Do not plan local drawing, deterministic text rendering, overlays, screenshot edits, or manual bitmap painting as a fallback for content/text. If generated text is wrong, unreadable, or malformed, the asset must be regenerated through the image model. Local resize/crop/pad is allowed only as platform-size normalization after generation, with no text/content changes and with a recorded review gate.

When to route here vs straight to cmm-image-creator

Mirror of the video line's planner/executor split (see cmm-video-production):

• cmm-visual-production = decide & spec. Use it when there is a real visual decision: choosing layout/format, building a card-set or cover system, or when a reviewable visual_package is wanted before generating.
• cmm-image-creator = generate. When the output form is already fixed — a single image at a known size, a one-off cover, or reusing/regenerating a finished card — the agent may call cmm-image-creator directly; it accepts either a visual_plan spec or a direct prompt.

Don't force the extra hop through this layer for an obvious single-image job. When this layer runs, it hands cmm-image-creator the asset spec to consume.

Input

Preferred input is a content_package from cmm-content-writing:

content_package:
  plan_id: ""
  platform: ""
  format: ""
  title: ""
  body_markdown: ""
  summary: ""
  variants: []
  sources_used: []
  visual_needs: []
  video_needs: []
  publish_notes: ""
  fact_check_status: pending
  status: draft

If visual_needs is empty, infer only the minimum useful assets from platform and format. Do not overproduce images.

For WeChat article body images, no count recommendation may be made before the body-image count gate is computed. First report the article body character count, the matching row from article-illustration-method.md, the minimum required body images, the recommended count, and whether the proposed plan needs a waiver. Only then choose image jobs and placements. A compact article, a strong single workflow, or an aesthetic preference is not a valid reason to skip this gate.

Read references/visual-package-contract.md for exact fields. Read references/visual-method.md for platform asset rules, prompt structure, image-generation routing, and quality checks. Read references/article-illustration-method.md whenever planning body illustrations for an article (功能分类、位置触发器、来源路由、illustration_map)。 Read references/platform-image-specs.md for exact per-platform dimensions and references/wechat-visual-style.md plus templates/ for WeChat article layout and body-illustration patterns.

Before building a visual_plan, resolve the upstream content_plan.style_selection: load the visual template from cmm-content-factory/references/styles/visual/<id>.yaml and the layout template from cmm-content-factory/references/styles/layout/<id>.yaml. Their palette, fonts, image tone, generation avoid-lists, layout text/table/image rules, and consistency rules are mandatory constraints for asset specs, prompts, and article layout. Visual and layout temperature must match per styles/selection-rules.md.

Workflow

1. Validate the content_package.
2. Determine platform requirements:
   • aspect ratio
   • number of assets
   • cover vs body illustration vs card set
   • whether screenshots or user-provided images are required
3. Build a visual_plan.
4. For card sets, build a card_text_map that binds each card to title, body copy, source section, and caption role.
5. Build asset specs, rendering policy, post-render checks, and final image prompts.
6. Ask for confirmation before invoking any image-generation or rendering skill.
7. Route to an approved backend only after confirmation.
8. Preserve prompt files and asset paths in visual_package.
9. Run or record quality checks.
10. Stop before publishing.

Routing

| Need | CMM output | Execution after approval | |---|---| | WeChat cover pair, editorial social cards, polished carousel, screenshot-heavy post | wechat_cover_pair or social_card_set | built-in card layout + cmm-image-creator | | Single article/blog/concept cover image | cover | cmm-image-creator | | Article body illustrations, diagrams, comparisons, flowcharts, infographics | article_illustrations | built-in illustration method + cmm-image-creator | | Xiaohongshu/Rednote image card series, knowledge cards, checklist cards | xhs_card_set | built-in card layout + cmm-image-creator | | Raw image generation (any runtime) | generated_image | cmm-image-creator (prefers native runtime image generation, falls back to configured third-party APIs after approval) | | Prompt-only planning | image_prompts_only | none |

Routing rules:

• Use CMM asset specs for layout, typography, platform ratios, and visual packaging first.
• Use the built-in card layout method when layout, typography, platform ratios, or visual packaging matter more than raw image generation.
• Generate a standalone raster cover via cmm-image-creator.
• Use the built-in illustration method when the article body needs explanatory visuals tied to sections.
• Build a card series (social browsing) via the built-in card layout + cmm-image-creator.
• For Xiaohongshu/Rednote text-heavy cards, the final card image content/text must also come from the approved image-generation backend. Do not switch to local PIL/HTML rendering or manual bitmap composition for text/content. Platform-size normalization may use local resize/crop/pad only when it preserves the card content and is recorded for review.
• Model selection is not decided here. Hand cmm-image-creator the asset spec only — prompt, size/aspect-ratio, resolution (1K/2K/4K), role, and any reference image. Do not pass a provider/model from this layer; cmm-image-creator owns model scheduling: host-native first (Codex native image generation in Codex), then third-party fallback only when needed. The optional_backend / provider_policy fields in visual_plan should be left empty (or "auto") unless the user explicitly pins a provider for a specific asset.

visual_plan

Before generation, output a plan:

visual_plan:
  content_package_id: ""
  platform: ""
  required_assets:
    - asset_type: ""
      count: 1
      aspect_ratio: ""
      purpose: ""
      asset_spec:
        canvas: ""
        text_zones: []
        image_zones: []
        typography: ""
        color_direction: ""
      prompt: ""
      optional_backend: ""
      provider_policy: ""
      confirmation_required: true
  illustration_map: []   # 正文配图逐图登记,见 references/article-illustration-method.md
  card_text_map: []
  rendering_policy:
    preferred_renderer: ""
    text_rendering: ""
    generated_image_role: ""
  post_render_checks:
    required: true
    checks:
      - dimensions
      - minimum_font_size
      - text_overflow
      - overlap
      - safe_margin
      - mobile_readability
      - module_boundary
      - corner_marker
      - left_rail_absent
      - checklist_visibility
      - tag_inset
      - text_marker_alignment
      - design_quality
  source_images_needed: []
  style_direction: ""
  layout_notes: ""
  risks: []
  quality_gate:
    readability: ""
    platform_ratio: ""
    source_integrity: ""
    text_overflow: ""
    brand_safety: ""
    design_quality: ""
  status: pending_confirmation

visual_package

After confirmed production, output:

visual_package:
  content_package_id: ""
  platform: ""
  asset_type: cover
  visual_style:
    id: ""
    cover_source: ""
    illustration_source: ""
  body_image_count_gate:
    article_chars: 0
    article_type: ""
    minimum_required: 0
    recommended_count: 0
    planned_count: 0
    count_status: pass
    rationale: ""
    replacement_breakpoints: []
  body_image_count_waiver:
    required_count: 0
    planned_count: 0
    missing_count: 0
    waiver_reason: ""
    replacement_breakpoints: []
    reader_impact_assessment: ""
    user_confirmation_required: true
    user_confirmed: false
  image_text_lock:
    article_title: ""
    body_final_status: ""
    cover_text:
      title: ""
      subtitle: ""
    body_images: []
  asset_paths: []
  prompt_paths: []
  image_prompts: []
  layout_notes: ""
  style_notes: ""
  source_assets: []
  quality_notes: []
  post_render_report:
    cards:
      - card: 1
        path: ""
        dimensions: ""
        min_font_px: 0
        body_font_px: 0
        title_font_px: 0
        text_overflow: ""
        overlap: ""
        safe_margin_px: 0
        mobile_readability: ""
        module_boundary: ""
        corner_marker: ""
        left_rail_absent: ""
        checklist_visibility: ""
        tag_inset: ""
        text_marker_alignment: ""
        design_quality: ""
        needs_manual_review: false
    summary: ""
  quality_gate:
    readability: ""
    platform_ratio: ""
    source_integrity: ""
    text_overflow: ""
    brand_safety: ""
    design_quality: ""
  status: draft

Use body_image_count_gate for new WeChat tasks. body_image_count_rationale is deprecated and must not be used as the only count record. If planned body images are below the minimum from article-illustration-method.md, add body_image_count_waiver and keep user_confirmed: false until the user explicitly confirms the waiver in chat. Do not present an under-minimum plan as the default visual plan; present it only as a waiver option next to the compliant plan.

Quality Gate

Gate	Requirement
readability	Text is readable at mobile size.
platform_ratio	Aspect ratio and safe zones match target platform.
source_integrity	Visualized products, people, facts, and screenshots are sourced or marked for review.
text_overflow	No card/cover text overflows or collides with other elements.
brand_safety	Avoid misleading logos, unsupported claims, or unapproved impersonation.
design_quality	Visual hierarchy, spacing, information density, and card-to-card consistency are good enough for the target platform.
style_consistency	Assets follow the selected visual and layout templates: palette, fonts, cover template family, ratio per platform, and model-rendered text requirements.

Confirmation Rules

• Default behavior is plan first, generate second.
• Do not call image generation, Playwright rendering, or screenshot export until the user confirms the visual_plan.
• For WeChat article runs, the confirmation must include the final title, final body status, per-image text lock, cover text, and body-image placement/order. A broad instruction such as "发到草稿箱" is not enough to approve unseen image text or layout.
• If the user explicitly says "直接生成", "不用确认", "跳过确认", or equivalent, state the assumptions before generation. This shortcut cannot bypass the WeChat cover-text requirement or a stricter upstream factory preview gate.
• Preserve prompt files whenever the routed skill requires them.
• If a routed legacy skill has its own stricter confirmation rule, follow the stricter rule.

Stop Conditions

Stop and ask before:

• Generating paid image assets.
• Fetching web images for current products or people.
• Publishing/uploading images externally.
• Replacing existing visual assets.