Imported from NimaChu/xhs-imagen (SKILL.md). Install upstream with npx skills add NimaChu/xhs-imagen. Copyright stays with the author.

xhs-imagen

Turn one topic or source article into a publication-ready Xiaohongshu post package.

Required result

Create:

output/<topic-slug>/
├── project.json
├── research.md
├── post.md
├── storyboard.md
├── prompts/
│   ├── 00-cover.md
│   ├── 01-*.md
│   └── ...
├── images/
│   ├── cover.png
│   ├── page-01.png
│   └── ...
└── qa-report.md

If the user requests only part of the package, create only that part. Never claim that images were generated when no image-generation or local SVG renderer was available.

Fixed Xiaohongshu defaults

• Create a 3:4 cover, recommended 1080 × 1440.
• Create 9:16 information pages, recommended 1080 × 1920.
• Default to one cover plus 5–8 information pages.
• Use Simplified Chinese unless requested otherwise.
• Explain one dominant idea per page.
• Optimize titles and labels for phone reading.
• Do not render page numbers or page-position markers such as 04, 01/08, or PAGE 04. Keep ordering only in filenames. Numbered steps are allowed when the numbers explain the content itself.
• Default to the alpaca-line-art profile: pure-white background, fine black hand-drawn lines, and the bundled white alpaca creator IP.
• Preserve glasses-chibi-blue as a selectable profile for the original glasses-wearing host, warm off-white paper, and cobalt-blue comic style.
• Use toolbox-bot-risograph for tool ecosystems, plugins, Skills, Agents, and workflows in two-color risograph.
• Use maker-girl-editorial for professional AI Coding, workplace, tutorial, and opinion content in modern editorial illustration.
• Use cyber-luban-woodcut for Skill–Harness–Agent architecture and system-building topics in new-Chinese woodcut.
• Use capybara-gouache for beginner explainers, pitfalls, reassurance, and everyday analogies in warm gouache.
• Make the character perform the page's core conceptual action; never use it as corner decoration.

Read references/visual-profiles.md, references/visual-style.md, and references/character-consistency.md before producing images.

Workflow

1. Resolve the brief

Determine or infer:

• topic, audience, and desired outcome;
• the single sentence readers should remember;
• source material and whether facts may have changed;
• page count and language;
• selected visual profile, character, palette, and brand constraints;
• whether the user wants a complete package, images, copy, or prompts.

Use beginner-friendly AI/technology education as the default audience and tone when the request does not specify them.

Store the choice in project.json as visual_profile. Honor an explicit user choice; otherwise use the default declared in references/visual-profiles.json. Use one profile for the whole series unless the user explicitly requests otherwise.

When reviewing pages the user selected or rejected, distinguish explicit feedback from inferred preference. Treat only explicitly confirmed rules as durable defaults; use the final selection primarily to understand visual appeal and expression accuracy rather than infer rigid layout rules.

2. Research before writing

Search primary and authoritative sources for current, technical, disputed, product-specific, numerical, legal, or attributed claims. Write a claim table to research.md. Separate sourced facts from analogies and editorial framing.

Read references/fact-checking.md for detailed rules.

3. Build the content arc

Create:

• one thesis;
• one useful analogy;
• one misconception;
• 4–7 supporting ideas;
• one limitation, boundary, or human-control point;
• one final takeaway.

Select pages by cognitive anchors instead of distributing content evenly. Keep only moments that change what the reader understands: a core judgment, cognitive turn, comparison, bottleneck, boundary, common mistake, state change, or takeaway. Drop a page when removing it does not weaken the learning arc.

For every selected page:

1. state the cognitive anchor and why it deserves a page;
2. convert the abstract concept into a physical action;
3. map that action to one ordinary low-tech object;
4. make the character perform the action so the metaphor depends on the character.

Write post.md with a Xiaohongshu title, publishable body copy, optional source note, and relevant hashtags. Write storyboard.md before generating images.

Read references/content-planning.md when choosing pages and reducing copy. Read references/visual-metaphors.md before writing the storyboard or image prompts.

4. Create and validate the project

Store exact content and page decisions in project.json. Start from references/project.template.json and follow references/project.schema.json.

Validate it:

python3 scripts/validate_project.py /absolute/path/project.json

Generate the image-model prompt files:

python3 scripts/make_prompt_pack.py \
  /absolute/path/project.json \
  --output-dir /absolute/path/output/prompts

5. Choose the rendering path automatically

When an image-generation tool is available

Use it as the default path.

1. Resolve the selected profile in references/visual-profiles.json.
2. Use that profile's character_reference as the only bundled image reference. Never attach multiple profile references to one generation call.
3. Generate the cover and one representative inner page first.
4. Inspect character identity, core action, metaphor originality, typography, spacing, color, copy accuracy, and absence of page-position markers.
5. Lock the successful visual description.
6. Generate the remaining pages using the same reference and style lock.
7. Save files as cover.png, page-01.png, and so on.

Do not invoke the local renderer merely to pre-empt possible text errors.

When no image-generation tool is available

Use the bundled local SVG-to-PNG renderer:

python3 scripts/render_xiaohongshu_project.py \
  /absolute/path/project.json \
  --output-dir /absolute/path/output

This path preserves the Xiaohongshu cover and inner-page ratios while converting the project into deterministic local knowledge-card layouts. Read references/local-rendering.md for limitations and renderer requirements.

6. Repair text only after explicit user feedback

Do not create a separate hybrid workflow. If the user explicitly identifies incorrect, corrupted, or unreadable text in an existing image:

1. Confirm the target image, exact replacement text, and affected region.
2. Prefer local image editing or regeneration when available.
3. If the problem remains, apply a deterministic SVG overlay only to that region:

python3 scripts/patch_image_text.py \
  --input /absolute/path/page.png \
  --output /absolute/path/page-fixed.png \
  --visual-profile <selected-profile> \
  --x 100 --y 300 --width 880 --height 180 \
  --text "正确文字"

4. Inspect the repaired image before delivery.

Never apply an SVG text patch speculatively.

7. Inspect every output

Write qa-report.md. For every failed check, record the defect, repair action, and recheck result; do not stop at listing problems. Verify:

• correct ratio and orientation;
• readable, accurate Chinese and product names;
• background, line treatment, palette, and typography match the selected visual profile;
• one dominant idea per page;
• a meaningful cognitive anchor on every page;
• an original physical metaphor with one primary structure;
• the character performs the metaphor's core action;
• stable profile-specific character identity and proportions;
• valid diagram flow;
• no cropped titles, faces, hands, or summaries;
• no unsupported factual claims or invented quotations;
• no visible page number, page count, or page-position marker; content-level numbered steps remain allowed;
• ordered, stable filenames.

Run:

python3 scripts/check_png_ratios.py /absolute/path/output/images

Read references/quality-checklist.md for the full review.

Final response

Provide:

• a concise summary of the content arc;
• the cover and ordered pages, or links to their files;
• the publishable post copy;
• source citations for time-sensitive claims;
• an honest note about any unresolved image or text defect.