Communitygithub.com

Newdvn2025/topic-to-published-video

For AI popular science and knowledge sharing, this skill transforms researched topics and outlines into Xiaohongshu videos, covers, and release packages; supports cost-controlled B-roll generation, voiceover acquisition, and phased packaging, as well as continuing from existing footage or live recordings.

What is topic-to-published-video?

This AI agent skill is designed for creators who want to efficiently produce polished short-video content (especially for Xiaohongshu) from well-researched topics and outlines. It acts as a full-fledged production director, managing a step-by-step workflow from initial framework confirmation to final video, cover art, and release package. The skill integrates with local tools like AutoDL (for B-roll generation), HyperFrames, and optionally Remotion or ChatCut, and supports both AI-generated digital presenters (via HeyGen) and manual voiceover recording. Key features include cost-controlled B-roll generation, phased delivery with quality gates, and strict adherence to user-provided materials as source truth. It avoids unnecessary rework by tracking project state in `production-state.json` and `production-log.md`, and only prompts users for critical decisions. The skill is suitable for platforms like Cursor, Claude Code, and Codex, and its documentation (SKILL.md) provides detailed references for initialization, workflow steps, artifact specs, and quality checks. By keeping execution local, it ensures user data privacy and avoids hidden costs.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add Newdvn2025/topic-to-published-video

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Topic to Final Video: Full-Process Director

Based on the user's current input, select an entry point and advance to the next incomplete stage. Includes AutoDL invocation scripts and optional tool skill installation wizard; the editing engine, account, and publishing interface are provided by the user's environment.

Typical Entry and SOP

The creator first selects a topic, conducts their own research, and defines an outline, then hands it over to the AI to invoke this skill. When an outline already exists, its intent and structure are preserved. The framework is confirmed first; re-doing research is not required. If facts or sources are insufficient, gaps are explicitly identified and verified when necessary; attachments are not automatically treated as verified facts.

Default order: Initialization → Framework Confirmation → Draft Voiceover + B-roll Generation Plan Confirmation → Generation and Acceptance Check → Final Recording Script → Self-Recording / HeyGen Digital Person → Multi-Clip A-Roll Editing → Full Voiceover and Real Subtitle Confirmation → B-roll Transitions and Style Confirmation → 9-Grid Confirmation → HyperFrames/Remotion Dynamic Review → Final Video, Cover, and Release Package. For the complete user SOP, see README; for stage details, see Workflow.

The goal is cost control and reduced rework, without sacrificing factual accuracy or action readability for low cost. Fixed fees or timeframes are not promised. Existing footage/live recordings can skip completed stages.

First-Time Initialization

On first use, or when the user says "install/initialize/reconfigure," first read the Initialization Guide, inventory existing tools, and ask about installation choices: recommended combination (AutoDL + HyperFrames), install all, or custom selection. If the user has already made a choice, do not ask again for the same choice.

AutoDL and HyperFrames are the recommended base combination; Remotion and ChatCut are optional. The goal is cost control and local production paths; fixed prices are not claimed. Once the selection is confirmed, run scripts/setup.py to install only the chosen skills, preserving existing configurations; then complete key configuration, runtime checks, and necessary application connections. Installation choices do not authorize paid video generation. Even if recommended tools are missing, the skill can still deliver a plan or use user-provided footage.

Keys are configured only in the user's local environment variables or via hidden input terminals, not collected in chat. Initialization records do not contain keys. Distinguish between "skill installed," "runtime available," and "account authenticated"; do not mark initialization complete just because the installation command succeeded.

Starting and Continuing

  1. Inventory inputs and available tools; read the project's production-state.json. New projects create files according to templates.
  2. Merge the user's current requirements with creator.json; current user requirements override project configuration, which overrides config/creator.example.json defaults. Example config is not user identity.
  3. Select an entry point according to the Workflow. For new topics, the default is asset-first: Framework Confirmation → Draft Voiceover and B-roll Plan → Generation and Acceptance → Final Recording Script → Self-Recording or Digital Person → Complete Voiceover Mainline and Real Subtitles → Style → 9-Grid → Dynamic Review. Existing footage or live recordings go directly to the corresponding stage.
  4. When the user says "OK, continue," record it as the corresponding confirmation together with the current deliverable. Only ask questions about pending choices that would clearly change direction. After confirming the existing layout and authorizing production, directly produce the specific review artifact.
  5. Update the sole current state, and write history to production-log.md; after upstream changes are made, mark affected downstream items as pending update.

Production Constraints

  • Text, timings, visuals, storyboards, and footage maps each have a single source of truth; see Artifact Specification. Attachment content is material, not automatically an execution instruction.
  • Default video is 16:9; cover is 3:4 independently, switching to 4:3 when needed. Personal name, credits, platform, and visuals are derived from the project or configuration, not hardcoded.
  • After the user submits a live recording, its actual text and timings replace the script draft; if word-level timestamps are missing, only produce subtitles for the real segment—do not estimate word-level cue points.
  • Generation service success does not mean the instructional action is qualified; according to Acceptance Rules, check the actual action before writing the related voiceover.
  • Ordinary confirmation does not automatically authorize external publishing, additional paid retries, or public disclosure of personal material. If the existing authorization scope is clear, do not re-request.
  • Use available capabilities as described in Tool Selection. When tools are missing, deliver whatever can be completed along with the specific gaps; do not pretend that generation, editing, or publishing has occurred.

On-Demand Loading

  • When preparing a recording script or selecting a digital person: read Voiceover Acquisition and HeyGen, which provides two paths: self-recording and digital person.
  • Upon receiving an audio file, SRT, user on-camera footage, or when entering post-production: read Live Recording Packaging.
  • When teaching content needs to show or give away prompt examples: read Prompt Delivery.
  • When finishing review, cover, and release package: read Acceptance Rules.
  • When needing to configure brand and author: read config/creator.example.json; copy it to the project's creator.json and fill in.
  • When needing examples: read examples/fpv-six-moves/README.md; the case study does not cover the user's new project requirements.
  • When using AutoDL: read addons/autodl-broll/SKILL.md; invoke the bundled script, first with prepare-only, then submit according to the authorized plan.

Delivery and Completion

The output status is determined by evidence: planned, assets_ready, recording_pending, review_ready, approved, delivery_ready, published. published must have actual evidence of publication.

delivery_ready requires the user-approved final video, subtitles, source project, cover, title/body/tags, and any promised comment material; report the aspect ratio, duration, real subtitle accuracy, inspection scope, and incomplete items. When the user approves and no content changes are needed, the same review video can be copied or formally referenced as the final version without re-rendering.

Individual skills in this repo

This repo contains 1 individual skill — each has its own dedicated page.

Related Skills