QC Image Series

Purpose

Turn a source object into a controlled image-series spec. The output should feel visually consistent and make the source's core information immediately legible. For chapter dividers, produce fixed-layout title-card prompts only by default; render actual divider images only when the user explicitly asks for raster assets.

Deliverable

Create exactly one short slug folder under qc-image-series-output/ in the current repository root. Put exactly one file inside it: series-spec.md.

Core Rules

• Message first: every image needs one concrete claim and one visual proof/cue. If the selected style uses text, add exact visible text; if the style is textless, set Visible text: none.
• Text policy comes from the selected style. Never plan a separate base image followed by Python, canvas, SVG, Pillow, ImageMagick, or other post-processing text overlay.
• Keep platform, layout, and style separate: platform decides ratio, safe zones, and text scale; layout decides persuasion structure and reading path; style decides palette, geometry, and visual language.
• No-frame flat field: no frames, cards, panels, windows, browser chrome, app mockups, terminal windows, screenshots, slides, dashboards, dialogs, modals, bordered layouts, pills, badges, or tabs. Put content directly on a clean background. Rectangles may be symbols, blocks, marks, or fragments, never enclosing containers.
• Chapter dividers are a separate mode: use the same fixed template reference in every prompt, change only the chapter title text, and do not add focal objects, evidence cues, icons, decoration, or layout variation.

Fixed SOP

Follow in order. Do not reorder.

1. Folder: create one short slug folder under qc-image-series-output/ in the current repository root.
2. Source: if media/audio/URL/subtitle/transcript, read references/media-source.md and follow its Decision SOP; otherwise use source text directly.
3. Platform: if user names 小红书/抖音/bilibili/B站 or multiple platforms, read references/platforms.md. If only 抖音, use douyin-vertical. If none, use base 16:9.
4. Mode: if the user asks for chapter split images, section divider images, chapter title cards, or other divider-only visuals, read references/chapter-divider.md, follow it, and skip style selection and layout selection.
5. Style: read references/style-selection.md. List styles from styles/*.md; if the user did not name one, ask them to choose before continuing.
6. Message units: each image gets core message, visible text or none, focal object, evidence cue, reader takeaway. Keep source-backed numbers/names/contrasts/timelines/causes/consequences/tensions. Merge or drop weak units.
7. Layout: read references/layout.md. Apply its Decision SOP top to bottom and stop at the first matching pattern.
8. Prompt: repeat platform + layout + style tokens in every prompt. Change only image-specific message fields. Follow the selected text policy inside the prompt.
9. Write: create only series-spec.md in the output folder.
10. QA: run the checklist before finishing.

Prompt Assembly Order

Use this order for normal message-series prompts. For chapter divider prompts, use references/chapter-divider.md. Timed or motion beats become static Scene state.

Platform: [platform], canvas [ratio], safe zone [margin].
Layout: [pattern]. Reading path: [first -> second -> third].
Style: [paste the selected style prompt block].
Core message: [one source-backed claim].
Text instruction: [render exact visible text when in-image-text; Visible text: none and forbid readable text when textless].
Focal object: [one object].
Evidence cue: [number, contrast, timeline, name, or consequence].
Scene state: [static decisive state].
Final image must already follow the selected text policy. Do not add text later.

Output Shape

Write series-spec.md in this shape:

# <Source Object Name> · Series Image Spec
Platform: xiaohongshu | douyin-vertical | douyin-horizontal | bilibili | base
Canvas: ...
Mode: message-series | chapter-divider
Style: <selected style id or chapter-divider fixed reference>
Text policy: in-image-text | textless | chapter-title-only
Series tokens: platform + layout + style tokens | fixed chapter divider tokens
Source: ...
Source normalization: existing transcript | sidecar subtitles | downloaded subtitles | embedded subtitles | ASR transcript | direct text
## Message Map
- ...
1. Image title
   Core message: ...
   Layout: ...
   Text instruction: ...
   Prompt: ...
   Consistency notes: ...
## Consistency Check
- ...

Don't Do

• Do not generate images, media, HTML previews, contact sheets, thumbnails, scripts, rendering utilities, or hand-rolled media parsing unless the user explicitly asks for actual chapter divider raster assets.
• Do not propose generating a base image and adding text later with Python, canvas, SVG, Pillow, ImageMagick, design software, or any post-processing workflow.
• Do not create multiple deliverable files or support artifacts inside the output folder. Temporary media/transcript files may exist outside it while working; the final output folder contains only series-spec.md.
• Do not turn the source into generic posters or abstract symbols with weak source grounding.
• For chapter dividers, do not invent visual metaphors, objects, icons, evidence marks, subtitles, decorative accents, or per-image layout choices.

QA Checklist

• No content is enclosed in a card, frame, window, panel, or mockup.
• Media sources state how subtitles/transcript were obtained; do not invent transcript content.
• Platform ratio and style tokens are separate and both are explicit.
• Normal message-series images name one layout pattern and reading path, and every prompt follows the selected style text policy.
• Each image has one clear source-backed claim, a correct text-policy instruction, and enough clean background around it.
• Motion language from the source has been converted into static visual states.
• Chapter divider specs state the detected output language and repeat the same fixed layout reference in every prompt.
• If chapter dividers use a reference image, the spec records the user reference path or bundled default asset, role, preserved invariants, and known limits of image-generation layout locking.
• The output folder contains series-spec.md and no other files.