sunshineLixun/video-publish-skill

Prepare Video Publish

Complete the entire workflow in the current Codex task. Never start a review website, local HTTP server, background Agent, app-server, or separate Codex task.

Initial Workflow

1. Resolve the attached video and/or subtitle to absolute local paths.
2. Run scripts/video-publish.mjs prepare and retain the returned sessionPath in this task.
3. Read the complete transcript.
4. Read content-contract.md, write content.json beside the session, and run finalize until validation passes.
5. Add any user-supplied style or person images with references add.
6. Read cover-workflow.md, update the cover spec when references or user instructions require it, and generate four complete covers with Codex built-in ImageGen.
7. Register the generated files with covers add.
8. Present the publishing draft and all four local cover images directly in this task, then stop and wait for user feedback or explicit publishing confirmation.

Do not open publishing pages during the initial workflow.

Prepare Input

Replace $SKILL_DIR with the absolute directory containing this SKILL.md.

Video or subtitle:

node "$SKILL_DIR/scripts/video-publish.mjs" prepare --input "/path/input"

Video with a preferred subtitle:

node "$SKILL_DIR/scripts/video-publish.mjs" prepare \
  --input "/path/video.mp4" \
  --subtitle "/path/subtitle.srt"

Subtitle with a separately supplied video:

node "$SKILL_DIR/scripts/video-publish.mjs" prepare \
  --input "/path/subtitle.srt" \
  --video "/path/video.mp4"

The command returns sessionPath and transcriptPath as JSON.

Finalize Content

Ground every factual claim in the complete transcript. Never infer visual details from video frames.

node "$SKILL_DIR/scripts/video-publish.mjs" finalize \
  --session "/path/session.json" \
  --content "/path/content.json"

Conversation Checkpoint

After registering covers, respond in the current task with:

• Summary, publishing title, description, and exact cover copy.
• The current cover version and changed ratios.
• Four labeled local images using absolute Markdown image paths.
• The available platform names, clearly distinguishing the four Ego staging platforms.
• A concise request for either focused revisions or explicit confirmation with platform names.

Do not replace this checkpoint with a URL or ask the user to find another task.

Continue In This Task

For every later user message, continue from the same sessionPath:

• Title, description, or summary change: update content.json and rerun finalize.
• Cover copy change: update content.json, rerun finalize, update the cover spec, and regenerate the affected covers.
• Prompt or visual change: update the cover spec and run spec set.
• New reference image: run references add, then include its returned ID in the cover spec.
• Focused ratio change: regenerate only the named ratios and pass only those files to covers add; unchanged ratios inherit from the previous immutable version.
• General cover change: regenerate all four ratios.

Show the new version in this task after every cover revision.

Open Confirmed Platforms

Only proceed when the user explicitly confirms in the current task and names at least one platform. Do not treat an earlier platform preference or ambiguous approval as confirmation.

If the session has only subtitles, ask the user for the video and attach it first:

node "$SKILL_DIR/scripts/video-publish.mjs" video set \
  --session "/path/session.json" \
  --file "/path/video.mp4"

Then open only the confirmed platforms:

node "$SKILL_DIR/scripts/video-publish.mjs" platforms open \
  --session "/path/session.json" \
  --confirmed \
  --platform rednote \
  --platform bilibili

Platform IDs are rednote, bilibili, douyin, and wechat_channels. The runtime rejects opening without --confirmed, finalized content, a complete cover version, and a video.

After Chrome opens, stop. Do not fill forms, upload files, read cookies, inject scripts, click publish, or operate any publishing page.

Stage Confirmed Platforms With Ego Lite

Use this mode only when the user explicitly asks in the current task to automatically upload or fill publishing pages and names at least one supported platform. An earlier request to merely open pages is not authorization. Ego staging supports 小红书 (rednote)、抖音 (douyin)、Bilibili (bilibili) and 微信视频号 (wechat_channels). Ego Lite is the only staging runtime.

node "$SKILL_DIR/scripts/video-publish.mjs" platforms stage \
  --session "/path/session.json" \
  --confirmed \
  --platform rednote \
  --platform douyin \
  --platform bilibili \
  --platform wechat_channels

The command uses one persistent Ego task space per platform. It inspects in parallel, uploads selected platforms in parallel, serializes metadata and cover changes, verifies the live page in parallel, and hands the task spaces to the user for review. If a result reports AUTH_REQUIRED, tell the user to complete login in the handed-off Ego space. Resume the same command only after the user explicitly says to continue. Never ask for passwords, cookies, QR codes, or exported browser state.

Staging performs only these authorized actions:

• Upload the session video.
• Upload the current 3:4 cover to 小红书; upload 3:4 and, when available, 4:3 to 抖音; upload 4:3 to Bilibili; upload both 3:4 and 4:3 to 微信视频号. Use --skip-cover to keep all platform-generated covers. Bilibili and 微信视频号 require the session's native 4:3 asset when custom-cover upload is enabled.
• Fill platform-native fields: 小红书 title, prose, and real topic entities; 抖音 title, body, and real topic entities; Bilibili title, description, and exact tag chips; 微信视频号 title-led description with plain hashtags and an empty short-title field.
• Reuse only matching drafts. Bilibili alone may quarantine a foreign local draft by saving it and returning to a verified clean upload page.
• Wait for platform processing to finish, verify the final control without clicking it, persist evidence, and leave the task spaces open for review.

Do not add --confirm-original-rights unless the user truthfully confirms the current video may be declared original. It enables the verified 小红书、Bilibili and 微信视频号 declaration flows. Omitting it still stages the draft and leaves declarations unchanged. The vendored upstream Douyin adapter currently has no verified original-declaration gate or action; do not claim that it set one and do not guess selectors from the upstream README wording.

After the result reports review-required, summarize each platform result and any title adjustment or timeout. Stop without clicking publish. Do not solve CAPTCHAs, bypass risk controls, inspect cookies, export authentication state, or use another browser tool to complete the final action.

Runtime Capabilities

Do not require every optional tool before starting. Check capabilities only when the selected path needs them:

• Node.js 20 or newer for the bundled CLI.
• ffmpeg and ffprobe only for video input.
• Python 3 and faster-whisper only when a video has no usable subtitle track.
• Ego Lite and the ego-browser command only for staging.
• Google Chrome or Chromium only for opening confirmed publishing pages without staging.

The Skill installer does not silently install system software. On failure, read runtime.md, report the exact missing capability, and get user approval before making system-level changes.