Creating a Highlight Video

Overview

This is the master orchestration skill for the Highlight Editor Family. When the user pastes a video URL and asks for a highlight, you MUST immediately activate this skill. It acts as an automated pipeline director, ensuring the AI seamlessly hands off tasks to the correct sub-skills and scripts without requiring the user to manually invoke them.

When to Use

• Whenever the user asks to "make a highlight", "cut this video", or pastes a YouTube URL with the intent of extracting scenes.

The Automated Pipeline

[!IMPORTANT]
REQUIRED SUB-SKILL (FIRST): Before anything else, run preparing-tools to verify all system dependencies (yt-dlp, ffmpeg, python3, sqlite3). Do NOT proceed if any tool is missing.

When this skill is activated, you must guide the operation through these 4 strict phases:

Phase 1: Scene Selection

To cut a video, we need to know what to cut.

1. Ask the user if they already have the timestamps or a Markdown selection list.
2. If they do NOT have one, offer to read the transcript or use the livestream-scene-selection skill to help them identify the best timestamps (following the "But / Therefore" rule).
3. REQUIRED SUB-SKILL: Before finalizing any scene selection, you MUST apply references/no-jumpcut-rules.md. This skill enforces Walter Murch's Rule of Six, the 5-min buffer rule, the merge-contiguous rule, and the pacing rollercoaster — all of which prevent jump cuts in the final output.
4. Once you have the timeline validated against references/no-jumpcut-rules.md, write it out to a .md file (e.g., highlight_selection.md).

Workspace Rule: If there is an existing youtube<VIDEOID>_workspace directory (created by anibon-timestamper or similar skills), you MUST save this .md file inside that directory!

Phase 2: Highlight Planner

Once the .md file is ready, you must invoke the planner.

• Find [SKILLROOT]: Look at the <skill location="..."> XML tag at the top of your prompt. Strip the filename SKILL.md (or with backslashes \ on Windows). Replace all \ with /. The result is [SKILLROOT].

1. See highlight-planner/SKILL.md.
2. Run: python3 "[SKILLROOT]/scripts/planhighlight.py" <yourmarkdownfile.md> --video-id <ID> --source <URL>
3. Wait for the highlightplan<id>.json file to be generated.
4. CRITICAL DURATION CHECK: Read the stdout from the planner script. If the total highlight duration exceeds 30 minutes, STOP and warn the user. Tell them downloading will take a long time, and offer to list available resolutions/framerates using yt-dlp -F <URL>. If they choose a lower quality to save time, append --format "<their_choice>" during Phase 3.

Iron Rule: Never read the JSON file. It is for the machine only.

Phase 3: Highlight Cutter (Subagent Delegation)

Because video downloading and FFmpeg rendering can take time, this phase should ideally be handled cleanly.

1. See highlight-cutter/SKILL.md.
2. Run: python3 "[SKILLROOT]/scripts/cuthighlight.py" <yourplan.json> --source <URL> [--format "chosenformat"]
3. Monitor the background task or subagent. It will download the chunks (or use a local --source) and stitch them via FFmpeg filter_complex.

Tip: If the user already downloaded the full video to avoid YouTube rate limits, pass the local file path to --source instead of the URL!

Phase 4: Highlight Verifier

Once the .mp4 is rendered, we must QA check it.

1. See highlight-verifier/SKILL.md.
2. Run: python3 "[SKILLROOT]/scripts/verifyhighlight.py" <yourhighlight.mp4> <yourplan.json>
3. Read the output. If it PASSES, present the final MP4 file link to the user!

Zero-Friction Execution

The goal of this orchestrator is Zero-Friction. Do not stop and ask the user for permission between Phase 2, 3, and 4. Once Phase 1 is complete and the timestamps are approved, you should execute the Planner, Cutter, and Verifier in one continuous flow, only reporting back to the user when the final QA-checked MP4 is ready!

Subagent Dispatch Contract

1. Pre-Grant Workspace Permissions: Call sk_permission for

ead_file on working workspace directory.

1. Parallel Highlight Scene Selection: Spawn parallel subagents using invoke_subagent across transcript/video chunks to identify peak meme/hype highlights.
2. Assembly: Pass highlight timeline timestamps to dit-cut-video-ffmpeg for cutting and stitching.