XHS Feishu Delivery

Use this skill to operate the local workflow that creates a Xiaohongshu image-text content package and sends the complete package to Feishu for manual posting.

Safety Rules

• Do not call Xiaohongshu MCP, browser publishing scripts, or any publish API.
• Do not create, read, or upload Xiaohongshu cookies.
• Do not open Xiaohongshu editor pages or click publish buttons.
• Do not add Feishu buttons, callbacks, WebSocket receivers, or approval state machines.
• Do not include music fields.
• Do not treat this workflow as data analytics, Obsidian memory, competitor analysis, or self-improvement tracking.
• Do not replace the mature skill chain with ad hoc model guesses when a listed skill is available.
• Do not ask the user to choose image watermark, visual style, layout, palette, image backend, or preference save scope. This workflow already pins those choices in .baoyu-skills/baoyu-image-cards/EXTEND.md.

Required Mature Skill Chain

This skill is the orchestrator and delivery layer. It must reuse the mature
skills agreed for the owner's workflow instead of inventing substitutes.

Use this chain for a new post:

1. Topic / source selection:
   • Before selecting an automatic topic, inspect sent history:
     python scripts/run_xhs_delivery.py --workspace "<workspace>" --history
   • Avoid any topic that shares the same GitHub repo, source URL, topic_key, content_id, or review_id with sent history.
   • If the user gives a specific topic and source, use it.
   • If the user asks for an AI-circle topic, hot news, GitHub trend, major AI update, or gives no concrete topic, use aihot first to pull current AI news and choose one Xiaohongshu-suitable vertical topic.
   • If the chosen angle depends on GitHub stars, repo activity, or developer-platform facts, use agent-reach as an additional verification/research skill after aihot.
   • Do not invent a hot topic from general model knowledge when aihot is available.
2. Fact verification:
   • Use agent-reach for official sources, GitHub facts, repo activity, X posts, papers, or source URLs that materially affect the claim.
   • Do not rely on an aihot summary alone for concrete factual claims.
3. Content type and insight pack:
   • Classify the post directly as one of:
     github_project_recommendation, ai_product_release, ai_industry_shift, or ai_technical_breakthrough.
   • Write the insight_pack directly from the verified facts before writing the body. Do not invoke a separate analysis/report skill for this step.
   • The insight pack is a short planning object, not a visible template. It must contain the hook, event, importance, takeaways, use cases, actionable framework, source facts, boundaries, and reader payoff.
4. Title candidates: use dbs-xhs-title style rules for Xiaohongshu title formulas and candidates.
5. Body copy:
   • Read references/editor_prompt.md before writing body_full.
   • Use the accepted voice target in references/editor_prompt.md as the style anchor.
   • Write body_full directly with the current model from the verified facts and insight pack.
   • Do not invoke any separate copywriting rewrite skill. Extra copy layers can reintroduce template voice.
   • Before finalizing, privately rewrite once if the draft sounds like a tool-review template, checklist, product manual, or source recap.
   • The final body should sound like a real AI-tools Xiaohongshu creator sharing a useful discovery, not a report, listicle, product manual, or training handout.
6. Image-card structure and prompts: use baoyu-image-cards with the bundled xhs-warm-cute-open-source preference.
   • For GitHub stars, open-source projects, or repo-based topics, put the verified repo/project name, star count, license/open-source badge, and source cue on the first card. Do not hide these facts in later pages or invent unverified counts.
   • Use the bundled defaults non-interactively: no watermark, style xhs-warm-cute-open-source, layout balanced, palette macaron, backend imagegen, batch size 4, and confirmation skipped with --yes or the runtime's equivalent "use defaults directly" instruction.
   • If baoyu-image-cards asks first-use preference questions, answer them from the bundled config instead of asking the user.
7. Final PNG generation: use Codex imagegen or the user's equivalent real image model as the raster backend.
8. Packaging and Feishu delivery: use this xhs-feishu-delivery workflow.

If any required mature skill is unavailable in the current runtime, stop and
state exactly which skill is missing. Do not silently replace it with shell
scripts, browser screenshots, generic writing, or template rendering.

Workflow

Authoritative rule: this skill packages model-generated cards. If any workspace
note, old handoff, old script, or previous conversation says to create or run a
local image renderer, treat that instruction as obsolete.

1. Confirm or create the target workspace. For a new workspace, initialize it from the bundled template:
   python scripts/run_xhs_delivery.py --workspace "<workspace>" --init-workspace
2. The target workspace must contain the four workflow directories:
   asset-generation, image-generation, publish-mainline, and feishu-delivery.
3. Create asset-generation/content_spec.json from the current topic using the Required Mature Skill Chain above. The template intentionally does not ship a starter post, so do not reuse old workspace output or old examples as copy. The file must contain the current topic, content_type, insight_pack, title, body, tags, image slug, source verification, and exactly 6 image page definitions.
4. Check the current spec against sent history before image work:
   python scripts/run_xhs_delivery.py --workspace "<workspace>" --check-history
   If it reports duplicate, choose a new topic/angle/source. Only set history.allow_repeat: true when the user explicitly asks to repeat a topic.
5. Run the asset generator first. It writes the copy package and the six prompt files:
   python "<workspace>\\asset-generation\\generate_current_assets.py"
6. Generate the six image cards with the mature image-card path used by the owner:
   use baoyu-image-cards to structure the series and Codex imagegen as the raster backend.
   If those skills/tools are available in the current runtime, explicitly use them; do not replace them with shell, Python, browser, or canvas code.
   Invoke baoyu-image-cards with the workspace .baoyu-skills/baoyu-image-cards/EXTEND.md defaults and --yes / equivalent direct-default confirmation. Do not pause to ask about watermark text, style, layout, palette, backend, or preference save location.
   Save each final PNG exactly to the image_path listed in asset-generation/outputs/current-publish-assets.json.
   If the user has a different image model available, use that model only if it can save final PNGs to the same paths.
   If a real image model/backend is not available, stop and report that image generation is blocked. Do not create a Python, PIL, SVG, HTML, canvas, screenshot, browser, or placeholder renderer to work around it.
   If generated images are saved outside the workspace by the image tool, copy the selected final PNGs into the required image_path locations before packaging.
7. After all six model-generated PNG files exist, run the wrapper. It must refresh the asset package, build the local package, build the Feishu card, and then validate or send:
   python scripts/run_xhs_delivery.py --workspace "<workspace>" --local-only
8. After a machine restart, check Feishu credentials without building or sending:
   python scripts/run_xhs_delivery.py --workspace "<workspace>" --check-feishu
9. Before using the workflow from a new Codex window, run read-only diagnostics:
   python scripts/run_xhs_delivery.py --workspace "<workspace>" --doctor
10. Install a Windows logon health check when the workflow should recover after reboot:
    python scripts/run_xhs_delivery.py --workspace "<workspace>" --install-startup-check
    If Task Scheduler or the Startup folder is blocked by local permissions, the workspace installer may fall back to the current user's Windows Run registry entry.
    For "machine booted but no user has logged in", install the administrator-only SYSTEM startup task:
    python scripts/run_xhs_delivery.py --workspace "<workspace>" --install-system-startup-check
11. If Feishu credentials should be checked after building the package:
    python scripts/run_xhs_delivery.py --workspace "<workspace>" --dry-run
12. Only when the user explicitly wants delivery to Feishu:
    python scripts/run_xhs_delivery.py --workspace "<workspace>" --send
    A successful send must append or update content-history/sent-posts.jsonl with the sent title, topic, source keys, Feishu message_id, and send time.

Content Standards

• Topic must stay in the AI/tools/dev productivity vertical unless the user explicitly changes the niche.
• content_type must be one of github_project_recommendation, ai_product_release, ai_industry_shift, or ai_technical_breakthrough.
• insight_pack must exist before the body is written and must include:
  core_hook, one_sentence_event, why_it_matters, key_takeaways, use_cases, actionable_framework, source_facts, boundaries, and reader_payoff.
• The body must deliver at least one concrete method, formula, checklist, use case, or judgment framework. It must not be only a news recap or generic reminder.
• GitHub/open-source posts must default to a positive discovery/recommendation tone: explain what the project does well, where it is useful, who should save it, and why it is worth trying.
• AI hot-news posts must turn the event into a reusable judgment: what happened, why it matters, who it affects, and how the reader should evaluate similar events.
• The body must follow references/editor_prompt.md: no rigid "first/second/third" structure, no product-manual tone, no generic "worth watching" language, and no forced risk section for positive GitHub recommendations.
• Title must be 20 Chinese characters or fewer.
• Body must be 1000 characters or fewer.
• Tags must be a non-empty list.
• Image cards must be exactly 6 PNG files generated by an image model/skill.
• Do not create image cards with PIL, SVG, HTML, canvas, template drawing scripts, or placeholder diagrams.
• Do not create, restore, edit, or run any file named render_current_cards.py.
• Do not use screenshots, browser-rendered HTML, Mermaid, matplotlib, or presentation slides as final card substitutes.
• If a legacy local image renderer exists in a workspace, ignore it and remove it before delivery; it is not part of the accepted workflow.
• The Python wrapper is allowed to require the 6 final PNG files to exist because image generation is handled by baoyu-image-cards/imagegen, not by the packaging scripts.
• The wrapper must use the workspace lock .xhs_delivery.lock; do not run direct workflow and skill workflow concurrently in the same workspace.
• Feishu card must contain only: topic, title, full body, image list, 6 image previews, and tags.
• Cover cards for GitHub/open-source topics must expose the concrete project/source facts immediately while keeping the warm cute hand-drawn macaron style.
• Sent-history duplicate checks are mandatory before automatic topic selection and before final asset generation. Duplicate keys include normalized GitHub repo, source URL, topic_key, content_id, and review_id.

References

• For repository structure and every tracked file, read PROJECT_GUIDE.md.
• For the content spec shape, read references/content_spec.md.
• For the model-image handoff contract, read references/image_generation.md.

Validation

Before saying the workflow is ready, run:

python scripts/validate_skill_safety.py --skill-dir "<skill-dir>"
python scripts/run_xhs_delivery.py --workspace "<test-workspace>" --init-workspace
python scripts/smoke_test_skill.py --skill-dir "<skill-dir>"
python scripts/run_xhs_delivery.py --workspace "<workspace>" --check-feishu
python scripts/run_xhs_delivery.py --workspace "<workspace>" --doctor
python scripts/run_xhs_delivery.py --workspace "<workspace>" --history
python scripts/run_xhs_delivery.py --workspace "<workspace>" --check-history
python "<workspace>\asset-generation\generate_current_assets.py"
# Generate the 6 PNG files with baoyu-image-cards/imagegen, then:
python scripts/run_xhs_delivery.py --workspace "<workspace>" --local-only

For Feishu credential verification, run --dry-run. Do not run --send unless the user asks to send a real Feishu card.