Wenchang Orchestrator

Wenchang Orchestrator

Turn a content request into a staged workflow. Enter at the stage supported by the user's actual input and keep human judgment at fact, image-plan, external-write, and final-publish gates.

Core Workflow

source intake
  -> topic and research when needed
  -> evidence
  -> draft
  -> content review
  -> publish preflight
  -> platform branches
  -> local image production, adaptive generation, and QA
  -> image publishing plan when public URLs are needed
  -> publish final check
  -> manual publishing
  -> reusable asset capture

Do not force a reviewed draft through topic discovery or drafting again.

Platform Branches

After wenchang-publish-check preflight:

• WeChat or Zhihu long-form: route accepted image needs to article-to-illustrations, insert reviewed local images, then route illustrated Markdown to md-img-r2 plan mode when public URLs are required.
• Xiaohongshu, WeChat inline cards, Zhihu Idea, or generic carousel: route source material to long-to-cards with delivery_mode=publish_ready, then to cards-to-images. For generative or mixed work, let resilient-imagegen resolve generation_backend=auto; use chatgpt-image-handoff when Computer Use or manual handoff is selected. Continue until real local images pass per-image visual QA and a cards-manifest exists. Stop at human confirmation before platform upload.
• Explicit copy-only request: use long-to-cards with delivery_mode=copy_only and record that image production was intentionally skipped.
• WeChat Moments, community copy, or WeChat-specific insertion notes: use wechat-to-cards as an adapter around the shared card pipeline. Do not require it for ordinary long-form publishing.
• Public archive or URL reuse: route checked local images to md-img-r2 plan mode only when asset_url_policy=public. Direct platform upload can use local image files.

Merge the completed branch artifacts back into wenchang-publish-check final mode.

Automation And Gates

Proceed automatically when the next step is low risk:

• select the entry stage
• produce a research map, evidence pack, outline, or draft when inputs are sufficient
• run content review and publish preflight
• produce illustration decisions, card packages, approved-scope local card images, manifests, and dry-run image plans
• prepare a local ChatGPT Handoff Pack and inspect available generation capabilities
• update compact state and reusable case artifacts

Pause when:

• key claims lack evidence
• review recommends a major rewrite or platform change
• a long-form image plan, sensitive supplied asset, or paid generation needs new user judgment
• a Computer Use prompt submission or reference upload is the next action
• publish-ready card images have passed QA and await human confirmation
• an upload, URL rewrite, account action, or external publication is next
• current platform requirements or hot tags need live verification

Never treat a dry-run plan as an upload or a final package as permission to publish.

Compact State Contract

Keep full body text, publish copy, plans, and manifests in their own artifacts. Store pointers and statuses in content_state.

content_state:
  request:
    raw_intent:
    current_stage:
    target_platforms: []
  source:
    body_file:
    fact_status: unknown
    review_status: unknown
  platforms:
    wechat:
      body_file:
      publish_package_file:
      illustration_status: not_assessed
    zhihu:
      body_file:
      publish_package_file:
      illustration_status: not_assessed
    xiaohongshu:
      cards_file:
      caption_status: pending
      tag_freshness: unverified
  cards:
    delivery_mode: publish_ready
    platform_profile:
    package_file:
    images_dir:
    manifest_file:
    generation_backend: auto
    handoff_run_id:
    render_status: not_started
    visual_qa_status: pending
    human_confirmation: pending
    asset_url_policy: local
    r2_state: not_planned
  images:
    plan_file:
    manifest_file:
    generation_runtime_file:
    selected_backend:
    r2_report_file:
    r2_status: not_planned
  publish_gate:
    mode: preflight
    status: blocked
    blockers: []
  next_step:
    skill:
    reason:
    user_decision_needed:
  handoff:
    from_stage:
    to_stage:
    accepted_inputs: []
    ignored_context: []
    stop_condition:

Use these R2 states: not_planned, planned, confirmed, uploaded, and verified.

Use these card states:

• render_status: not_started, planned, generated, or failed
• visual_qa_status: pending, passed, or needs_revision
• human_confirmation: pending, approved, or rejected

For delivery_mode=publish_ready, a completed long-to-cards package with no generated images must route to cards-to-images. Never route directly from card copy to the final release decision.

Output Format

## Current Stage
- Entry:
- Active branch:
- Workflow:

## Completed
- ...

## Blockers And Human Decisions
- ...

## Recommendation
- Next skill:
- Reason:

## Artifact Pointers
- Source body:
- Platform bodies:
- Publish package:
- Card package:
- Card images and manifest:
- Illustration plan:
- Image manifest:
- Image-generation handoff:
- R2 report:

## content_state
```yaml

```

Boundaries

• Do not collapse the workflow into one oversized output or duplicate full bodies inside state.
• Do not treat research possibilities as verified facts or topic tags as verified hot tags.
• Do not let one skill silently overwrite another skill's approved artifact.
• Do not leave a publish_ready card branch at copy or prompt output.
• Do not upload, publish, delete, or perform irreversible account actions without explicit user confirmation.
• Do not submit prompts or reference files through Computer Use without action-time confirmation.
• Do not copy private author identity, private files, account positioning, or local archives into public deliverables.

作者入口

• 微信公众号：AI生命克劳德
• X：@yangchao228
• GitHub：yangchao228