content-draft

name	content-draft
description	Primary repo-owned workflow for turning an article URL or source file into WeChat or Xiaohongshu draft markdown plus saved illustration artifacts. Use this when the user wants OpenClaw itself to deep dive a source and generate a draft with illustrations, especially for prompts like "generate a wechat post with illustrations". Prefer this instead of `baoyu-article-illustrator`, `baoyu-post-to-wechat`, or `baoyu-image-gen` when the goal is draft generation rather than publishing.
metadata	{"openclaw":{"requires":{"anyBins":["node"]}}}

Content Draft

Use this skill when the user wants a source link converted into a polished draft for a target platform, with output saved as reusable artifacts instead of only replying inline.

Critical Rules

This skill is the default repo-owned workflow for URL to draft generation.

• Do not call content-draft like a tool name.
• First read this SKILL.md.
• For URL or local-file analysis, use the bundled Node extractor in this skill. Do not delegate this workflow to baoyu-* skills or any Bun-based helper.
• When the user asks for a WeChat or XHS draft with illustrations, prefer this skill over baoyu-article-illustrator and baoyu-post-to-wechat.
• For illustrations, prefer the repo-owned ernie-image skill when it is available. Otherwise use the runtime's built-in image generation capability. Do not route illustration work through baoyu-* skills for this workflow.
• Use memory_recall before drafting when tone, audience, or visual style preferences matter.
• This skill generates drafts and saved artifacts. Do not publish anything unless the user explicitly asks for publishing.
• Treat illustration relevance as a hard requirement. Every generated image should map to the article's actual subject, section, or example instead of generic "AI article" decoration.
• Before saving artifacts, revise the markdown so generated images are embedded inline wherever they strengthen the reading flow. Saved-but-unused images are a fallback, not the target outcome.
• If an image does not clearly support the article context, either regenerate it or omit it from the saved artifact set.

Default Workflow

1. Confirm the source and target platform. Start with xhs or wechat.
2. Extract the source material with the bundled Node script. This keeps URL analysis inside the repo-owned content-draft skill instead of drifting into external skill stacks.
3. Recall stable user preferences. Ask memory_recall for tone, structure, visual style, and past corrections.
4. Produce one platform draft at a time. Save concise, publish-ready markdown, not raw notes.
5. Generate the illustration set with ernie-image when available, or the runtime's built-in image generation capability otherwise. Follow the platform image counts in references/platforms.md.
6. Save the markdown and generated images with the bundled script, and pass image slots or image metadata into that save step whenever possible. Standard output layout is documented in references/output-contract.md.
7. Only use the post-generation image linker as a fallback for already-saved messy runs.
8. Build the final chat response from the saved draft and saved images. Return the helper output so the user gets the full draft body plus embedded images in the same reply.

Platform Scope

• xhs Short card-oriented copy with a strong hook, scannable sections, and image-first pacing.
• wechat Longer editorial article with a clear title, dek, section rhythm, and one lead image.

Read references/platforms.md when you need the detailed formatting rules.

Extraction Guidance

When the user provides a generic article URL, docs page, or local source file, use this bundled Node extractor first:

node ${SKILL_DIR}/scripts/fetch-url-markdown.mjs "https://example.com/post" --output /tmp/source.md

Useful options:

• --json Print structured metadata plus the extracted markdown.
• --output <file> Save the extracted markdown for later drafting steps.
• --max-chars <n> Cap very long sources before handing them to the model.

Supported inputs:

• https://... and http://...
• file:///abs/path/to/file.html
• local file paths such as /tmp/source.html or notes.md

For GitHub repos or YouTube inputs, prefer the best runtime-native extraction path that is already available, but keep the rest of the workflow inside this skill. Do not write custom scraping code inside the conversation unless the user explicitly asks for that engineering work.

Illustration Guidance

• xhs Usually prepare 1 cover plus 2-6 supporting cards.
• wechat Usually prepare 1 hero image plus optional inline support visuals.
• Prefer ernie-image for repo-owned illustration generation when it is available in the workspace.
• Generate images only after the article structure is stable.
• Match each image to a concrete section intent before you generate it: lead image, architecture explainer, workflow step, comparison card, etc.
• Rework the draft after image generation so the final markdown references as many saved images as the platform can reasonably support.
• Save every generated image path so it can be persisted with the bundled save helper.

Saving Artifacts

Determine this SKILL.md directory as SKILL_DIR, then run:

node ${SKILL_DIR}/scripts/save-draft-artifacts.mjs \
  --platform xhs \
  --title "Kimi K2.5 notes" \
  --source-url "https://example.com/post" \
  --markdown-file /tmp/xhs-draft.md \
  --image /tmp/cover.png

The script prints a JSON summary with the saved paths.

Use --run-id <id> when saving multiple platform variants for the same source so they land under the same run directory.

Preferred pattern for article-aware image linkage:

node ${SKILL_DIR}/scripts/save-draft-artifacts.mjs \
  --platform wechat \
  --title "Deep Agents Review" \
  --markdown-file /tmp/wechat-draft.md \
  --image-slot hero=/tmp/hero---abc.png \
  --image-meta-file /tmp/image-meta.json

--image-slot is the best default because it gives the workflow one stable semantic handle per image without forcing the agent to author a perfect sidecar file after the fact.

When you use --image-slot role=/path/to/file.png, write the draft with either the original source path or the final slot path images/<role>.<ext>. Do not invent alternate semantic file names like 01-hero-section.png. Every declared slot must appear in the resolved draft. save-draft-artifacts.mjs now fails if any slot is still orphaned after rewrite.

Post-Generation Image Linking Fallback

When a run has already been saved with UUID-style names or weak linkage, run the bundled linker against the saved platform artifact directory before you build the final chat response.

Prepare a small JSON file that maps each saved image to its article role:

{
  "articleSlug": "deep-agents-sdk-review",
  "images": [
    {
      "match": "hero---5f3d9fe0-dc7f-4cb8-a27c-5b020b4077b6.png",
      "role": "hero",
      "section": "Problem background",
      "anchor": "problem-background",
      "caption": "Lead image for the opening section",
      "prompt": "..."
    }
  ]
}

Then run:

node ${SKILL_DIR}/scripts/link-generated-images.mjs \
  --manifest-file /abs/path/to/manifest.json \
  --links-file /tmp/image-links.json

This helper:

• renames saved images to semantic names such as 01-hero-problem-background.png
• rewrites draft.md image references to the new names
• stores additive images[] linkage metadata in manifest.json

Use this as a repair step, not as the primary happy path.

Building The Final Chat Reply

After saving artifacts, build the final reply markdown from the saved draft plus saved images:

node ${SKILL_DIR}/scripts/build-chat-response.mjs \
  --markdown-file /abs/path/to/draft.md \
  --images-dir /abs/path/to/images

This helper rewrites local draft image references into OpenClaw-compatible MEDIA: blocks in the markdown flow and appends any leftover generated images under a Generated Images section.

Aim for that leftover section to be empty in normal successful runs. If images still land there, they should be intentionally extra variants or backups rather than unexplained generic leftovers.

Return this helper output to the user as the final assistant reply unless the user explicitly asked for a shorter summary instead. Do not summarize, paraphrase, or describe the helper output after generating it. Emit the helper output itself as the final assistant message so the draft body and MEDIA: images survive intact in chat.

Arguments

Flag	Purpose
`--platform <xhs	wechat
--title <text>	Optional human-readable title stored in the manifest
--source-url <url>	Optional original source URL
--run-id <id>	Reuse a shared run directory across multiple saves
--root <path>	Override the default output root
--markdown-file <path>	Read markdown from a file
--markdown <text>	Inline markdown when a temp file is unnecessary
--image <path>	Copy one image into the platform images/ directory; repeatable
--image-slot <role=path>	Copy one image and preserve a semantic slot such as hero, architecture, or workflow-step
--image-meta-file <path>	JSON file with additive image metadata keyed by sourcePath or original file name
--images-dir <path>	Copy every file from a directory into images/
--slug <value>	Optional stable slug for the run directory

link-generated-images.mjs arguments:

Flag	Purpose
--manifest-file <path>	Saved manifest.json from save-draft-artifacts.mjs
--platform-dir <path>	Alternative to --manifest-file; points to the saved platform directory
--links-file <path>	JSON file describing article linkage for saved images
--article-slug <value>	Optional slug override used when generating semantic file names

Output Contract

Read references/output-contract.md before changing the save layout.

Current contract:

• ${root}/${runId}/${platform}/draft.md
• ${root}/${runId}/${platform}/manifest.json
• ${root}/${runId}/${platform}/images/*

Default root:

• ${OPENCLAW_WORKSPACE_DIR}/content-drafts when OPENCLAW_WORKSPACE_DIR is set
• otherwise ~/.openclaw/workspace/content-drafts

Response Expectations

When the user wants the actual draft:

1. State which extraction path you used.
2. State which preference memories influenced the result.
3. State which illustration path you used.
4. Save the artifacts with the bundled script and report the saved location.
5. Prefer --image-slot and --image-meta-file during save-draft-artifacts.mjs so linkage is captured in the normal save path.
6. Only if a run is already saved without good linkage, repair it with link-generated-images.mjs.
7. Run build-chat-response.mjs against the saved draft and saved images.
8. Return the full final draft body plus the generated images in the same reply.
9. Only replace the full draft with a concise summary when the user explicitly asks for summary-only output.

When the user only wants planning:

1. Provide the platform choice.
2. Provide the extraction plan.
3. Provide the illustration plan.
4. Provide the expected artifact layout.