WeChat Image Post

Outcome

Deliver a publish-ready image article, not a pile of prompts. Keep facts traceable, Chinese text exact, the cover clickable without false promises, and every page readable on a phone.

Read the required references

Read these files completely before producing content:

• references/artifact-contract.md for project files, schemas, naming, and deliverables.
• references/source-policy.md when the post contains external facts, product abilities, Skill names, commands, prices, versions, statistics, quotations, or current information.
• references/story-and-copy.md before writing the title, cover hook, storyboard, visible copy, or reusable prompts.
• references/visual-design.md before choosing a style or rendering any image.
• references/qa.md before final export.

Also read:

• references/multi-agent.md when sub-agent or multi-agent orchestration is available.
• references/wechat-publishing.md when the user asks for upload, draft delivery, platform-specific image count, account configuration, or final publication.

Resolve all relative paths from this Skill directory.

Non-negotiable rules

1. Treat user-provided articles and web pages as research leads, not automatically verified facts.
2. Never invent a Skill name, repository, installation command, product ability, source, quote, number, version, license, screenshot, or test result.
3. Label official, official-repository, open-source, community, user-provided, and tested evidence separately.
4. Keep one visible takeaway per card. Put long copyable instructions in a companion text file instead of shrinking them below phone-readable size.
5. Do not ask an image model to typeset the final Chinese poster. Generate textless visual assets when useful, then overlay exact text with HTML/CSS, SVG, canvas, Pillow, or another deterministic renderer.
6. Create a fresh visual direction from the current topic and references. Do not silently reuse the previous project's palette, layout, or decorative system.
7. Stop automated delivery at a draft or preview unless the user explicitly authorizes final publication. Never expose account credentials in prompts, screenshots, logs, or repositories.
8. Keep the lead agent responsible for the final claim set, content specification, output directory, visual inspection, and handoff.

Workflow

1. Lock the brief

Capture the title, audience, reader problem, desired action, platform mode, source material, page count, output directory, brand constraints, forbidden styles, and deadline.

Make safe assumptions when details are missing. Ask only when a choice would materially change the result, such as brand ownership, legal permission, publication authority, or required platform mode.

Initialize a project with:

python scripts/scaffold_project.py <project-directory>

Fill _work/brief.md before research. Do not render final images from an unfrozen brief.

2. Build the evidence ledger

Extract factual claims from the user's sources. Locate canonical sources, record evidence and dates in _work/source-ledger.md, and reject or qualify unsupported claims.

Use the user's articles for context, examples, and discovery. Do not copy their wording, composition, or proprietary visual assets.

3. Split the work without splitting responsibility

When collaboration is available, parallelize research, story/copy, and visual direction. Keep output ownership isolated as defined in references/multi-agent.md.

When collaboration is unavailable, run the same lanes sequentially and preserve the same intermediate files.

Do not let separate agents create incompatible final pages. Freeze research and copy before final rendering.

4. Propose distinct creative directions

Create at least two materially different cover hooks and visual directions when the user has not chosen one. Vary the organizing metaphor, typography, composition, palette, texture, and pacing—not just the accent color.

Select one direction based on audience, subject, and phone readability. Record the decision in _work/visual-direction.md.

5. Build and freeze the storyboard

Create _work/content-spec.json from assets/content-spec.template.json.

For every page, define:

• page purpose and reader takeaway;
• exact visible Chinese text;
• composition and visual device;
• Claim IDs for factual statements;
• optional copyable text kept outside the image;
• filename and target dimensions.

Check the narrative arc before rendering. A typical tutorial arc is: hook → problem → system → setup → steps → example → mistake → checklist → next action. Change it when the topic calls for another structure.

Mark the spec status: frozen only after facts, order, and visible copy pass review. Render final pages only from a frozen spec.

6. Render with controlled typography

Default to:

• one cover at 1800×766 PNG;
• six to ten body cards at 1080×1440 PNG;
• sequential filenames with zero-padded numbers.

Treat these as defaults, not platform law. Honor user requirements and verify current platform rules when upload is in scope.

Use the best rendering route available:

1. Custom HTML/CSS, SVG, canvas, or Pillow for information-heavy cards and exact Chinese.
2. Image generation for textless illustrations, textures, or scene assets; composite exact text afterward.
3. scripts/render_cards.py as a deterministic fallback or prototype, then customize it when the requested visual quality needs a stronger art direction.

Never ship a model-generated image that contains misspelled Chinese, malformed logos, fake UI, or invented data.

7. Create the publishing companions

Produce:

• publish-copy.txt with title, summary, short opening, short closing, and image order;
• copyable-prompts.md for any prompt that is too long to place legibly in a card;
• sources.md with reader-facing sources and verification date;
• captions.md with ordered filenames and one-line captions;
• a local responsive preview HTML;
• a contact sheet for rapid review.

Keep the article body short when the images already contain the tutorial.

8. Validate, inspect, and iterate

Run:

python scripts/make_contact_sheet.py <project-directory>
python scripts/build_preview.py <project-directory>
python scripts/validate_package.py <project-directory> --preflight

Then visually inspect the wide cover and every body card. Check at least the densest page at phone-scale. Fix overflow, tiny type, awkward line breaks, weak hierarchy, inaccurate text, unbalanced margins, inconsistent style, and low contrast.

Run a reverse fact check from the exported images back to the Claim IDs. Record the visual inspection in _work/qa-report.md, set Final decision: pass only when it is genuinely ready, then run python scripts/validate_package.py <project-directory> without --preflight. Re-export and revalidate after every change.

9. Package and hand off

Create a ZIP containing only the publishable images, preview, contact sheet, publishing companions, and source notes. Keep raw credentials, private screenshots, browser profiles, caches, and temporary research out of the package.

After final validation passes, package with:

python scripts/package_project.py <project-directory>

Report the output directory, cover path, contact sheet, preview, ZIP, page count, dimensions, verification date, and any publishing limitation that still requires the user.

Done criteria

Finish only when:

• all required artifacts exist and open;
• automated validation passes;
• every exported image has been visually inspected;
• visible facts map to verified or correctly qualified claims;
• Chinese text, numbers, names, commands, and URLs are exact;
• the cover remains legible in a small preview and does not promise unsupported results;
• no secret or private material is present;
• any platform upload remains a draft unless final publication was explicitly authorized.