wang-bool/xiaohongshu-storycraft

Xiaohongshu StoryCraft

Convert user ideas, text, and images into publishable Xiaohongshu visual posts. Deliver complete title, body text, tags, per-page images, and source attribution — not just outlines or prompts.

Input & Boundaries

First, distinguish four types of material:

• Confirmed facts, data, and sources
• User-stated experiences, feelings, and judgments
• Content directly observable in images
• Interpretations and editorial expressions that can be inferred from the material

Do not fabricate experiences, effects, numbers, quotes, or sources that were not provided. When key conclusions lack evidence, ask only the minimal necessary questions; when the user requests direct creation, limit conclusions and clarify boundaries.

If the user provides no images at all, ask only once: A – Provide their own photos/screenshots (recommended by default); B – Use public image libraries; C – Use AI generation. Do not repeat after the user chooses. For factual evidence and processes, prioritize user materials or traceable public sources; pure AI images cannot serve as factual proof.

Do not auto-publish to Xiaohongshu. Automatic publishing requires separate platform connection, authorization, and publication confirmation.

Read References on Demand

• Before starting planning, read Content Quality Rules and Generative Layout Theory.
• When creating images, read Image Production Rules.
• When determining visual language, read Visual Constitutions; use only one constitution per set of images; after selection, continue reading only the detailed constitution for that style.
• When deriving or verifying color schemes, read Color Craft Principles.
• When using HTML, read HTML Production and Rendering.
• Only read Composition Failure Root Causes when composition repeatedly fails or requires post-mortem review. It is for diagnosis only, not as a source of alternative layouts.

Workflow

1. Extract facts, user experiences, image-visible content, and permissible supplementary explanations to form a factual boundary.
2. Plan the narrative for the entire set of images, writing a unique page proposition for each page. The number of pages is determined by the material and reader task; default to 5–7 pages, but do not repeat content just to meet a page count.
3. Read each real image and record its aspect ratio, focal points, information density, minimum readable size, completeness requirements, and narrative role.
4. Decide whether each page uses images or text to carry irreplaceable understanding and proof tasks. Establish narrative primary and visual weight based on this, deriving spatial relationships from content. Do not select, apply, or enumerate fixed templates.
5. Select or define a visual constitution. Determine it in order: user's explicit choice → user-described feel → content emotion, reader task, material color, and evidence density. Only ask when two choices would materially change the user's desired tone and the material cannot decide. The style influences trade-offs between effective compositions via color and spatial tone, but does not predefine the page skeleton.
6. Place real images and real text into HTML. First, scale the primary content to the size needed to complete its task, then observe remaining space to expand, compress, split, or reorganize text.
7. Verify image integrity, semantic line breaks, overflow, edge alignment, white space, hierarchy, and closure using a real browser measurement and full-page screenshots.
8. Record the composition fingerprint and compare it with adjacent pages. Re-derive if similarity lacks narrative justification; retain if there is a clear comparative or rhythmic role.
9. After completing checks for facts, sources, privacy, and public posting risks, deliver the output.

By default, complete judgments on behalf of the user. When materials are complex or page propositions have multiple substantive interpretations, first confirm the page-by-page outline; for routine tasks, generate directly and iterate internally.

Generation Boundaries

Components may only carry semantics such as title, body text, image, caption, number, quote, boundary, and source. Components must not predefine inter-component layouts.

Specific finished works and past cases can only explain why a composition succeeded or failed; they cannot be registered as selectable page templates, layout operators, or recipes for future generation.

The same source material may yield different compositions. Even for the same compositional relationship, dimensions and container sizes must be recalculated based on the real images and text; coordinates or container sizes cannot be copied.

Page & Image Minimums

Each page carries only one primary reading task and provides the context needed to understand that page's proposition. The narrative primary is determined by the irreplaceability of content, not by area, word count, or image aspect ratio.

Important images are by default kept intact with their subject, text, relationships, evidence, and necessary context; crop only when non-informational edges can be safely removed and cropping improves readability. Screenshots containing key text, buttons, data, or interface relationships cannot be redrawn as evidence by AI. Do not force an image on every page, nor force using all images.

When no source material can serve as a cover, AI-generated concept or atmosphere images may be used, but they must not impersonate real products, places, people, experiences, tests, or results. If there is any risk of misinterpretation, explicitly note it in the visual asset source.

Visual Generation & Editing Capabilities

• If the agent can see and generate/edit images: first read and preserve the original material's subject, text, relationships, and facts, then perform editing, integration, or constrained re-creation.
• If the agent can see images but cannot generate them: only perform cropping, color adjustment, collage, white space, text overlay, and privacy masking; when materials are insufficient, use text pages or indicate gaps.
• If the agent can neither see nor generate images: deliver only a visual plan, explicitly stating that images have not been read, edited, or verified, and cannot claim completed publication images.

When the user chooses public image libraries, download assets to the task material directory and record “filename ← source link” in SOURCES.md. Before finalizing, list sources and licensing risks; let the user decide how to attribute publicly.

Output & Rendering

Default canvas: 1080×1440 or proportional 3:4 vertical image. For fine typography, use assets/deck-template.html and scripts/render_deck.py. For simple collage, text overlay, or privacy masking, PIL may be used.

HTML pages use semantic markup and check hooks. Use data-phrase-lock for critical phrases, data-check-lines for text that must be checked at the last line, and data-preserve-ratio for images that must be fully displayed. See references/visual-style.md for specific conventions.

After rendering, run DOM checks and scripts/qa_deck.py, and review full-page screenshots and 360px thumbnails page by page. Automated measurements are for detecting issues, not replacing aesthetic judgment.

Delivery Format

Deliver in the following order:

## Published Final

### Title
[One final title]

### Body
[Body text ready for copy-and-paste publishing]

### Topic Tags
[#tags]

### Image Story
1. [Page name]
   - Page proposition [What the reader should understand or judge after viewing this page]
   - Image [Output file or explicit downgrade explanation]
   - Text in image [Key points on this page]
   - Factual basis [User material, public source, or no factual claim]
   - Visual asset source [User file, public asset, or AI generation]
   - Processing method/AI usage [Original image editing, integration, constrained re-creation, pure AI supplementation, or specific explanation]

### Publishing Notes
[Provide only when genuinely useful]

Pre-Publishing Checklist

• No fabricated facts, experiences, effects, data, or sources
• Title, body text, image text, and tags are consistent with each other
• Page propositions are clear; visual primary-secondary hierarchy between images and text aligns with that proposition
• Images that bear scene, action, result, evidence, or key relationships are not reduced to long-text illustrations; removing them would break page understanding or evidentiary chain
• Every piece of text provides new information that the image cannot convey alone; if page space is insufficient, check for proposition overload, image-text redundancy, and the possibility of splitting pages
• Images have been composed based on real aspect ratio, focal points, detail density, and minimum readable size, not pre-fitted into prefabricated containers
• Important images are displayed fully by default; completeness protects effective content and relationships, not mechanically preserving non-informational black borders; any cropping has a justifiable safe zone and narrative reason
• Formulas, number units, proper nouns, and short conclusions do not break lines in a way that destroys semantics
• Large areas of white space have a clear purpose and function; structural gaps have been resolved through content relationships or root causes of composition
• Sources, captions, and boundary notes are attached to corresponding content, not floating or pushed off the canvas
• Real screenshots have been checked page by page, not relying solely on source code or coverage metrics
• Composition fingerprints of adjacent pages have been compared; repeated compositions have a narrative rationale, not a generation habit
• The visual constitution achieves identity through color palette and spatial tone, without predefining page skeletons or forcing pages into repeated fixed structures
• Colors stay within the chosen constitution’s gamut boundaries; lightness steps, chroma order, and warm-cool tendencies between roles are valid; every color can explain its functional role; accent colors are scarce in area
• Base, surface, and border lines come from the same neutral color family; body text and de-emphasized information come from the same ink color family; sources, page numbers, and captions maintain ordinary small-text contrast against their actual background
• When side-by-side with the most confusable adjacent style, the gamut boundaries are immediately distinguishable; re-derived color values use constitutional anchors as gamut centers, constrained by lightness, chroma, color temperature, and role relationships
• When accent changes significantly in hue, re-validate the full color group, not just replace a single accent color; ultra-minimal styles like black-and-white that rely on neutrality continue to keep the base hue-free
• The entire deck uses only one signature language with a content reason; decorative candidates from the constitution are not accumulated as a checklist; the page works even without decorations
• Key factual images can be traced back to the user’s original image or a specific public source
• Pure AI images do not impersonate real content
• Checked for public risk of faces, contact information, ID cards, precise locations, and private screens
• All DOM QA and PNG QA checks have no FAIL; all REVIEW items have been verified via screenshots