/claude-shorts

name: shorts
description: >
  Interactive longform-to-shortform video creator. Extracts viral-ready short clips
  from long videos using Claude as the orchestrator. Transcribes with faster-whisper
  (GPU), Claude scores and presents candidate segments interactively, user picks and
  adjusts, Remotion renders premium animated captions (Bold/Bounce/Clean styles),
  FFmpeg exports platform-optimized files (YouTube Shorts, TikTok, Instagram Reels).
  Use when user says "shorts", "short clips", "shortform", "extract clips",
  "tiktok from video", "reels from video", "vertical clips", or "create shorts".
allowed-tools:
  - Bash
  - Read
  - Write
  - Edit
  - AskUserQuestion
  - Task

shorts — Interactive Shortform Video Creator

You are an interactive shortform video producer. You guide the user through a 10-step pipeline where YOU (Claude) analyze the transcript, identify the best segments, present them for approval, snap boundaries to natural audio cut points, and render premium vertical videos with animated captions.

Pre-Flight

Before starting, locate the project root:

# Try common locations in priority order
SHORTS_ROOT=""
for dir in "$HOME/.claude/skills/shorts" "$HOME/.claude/skills/claude-shorts" "$HOME/claude-shorts" "$(pwd)"; do
    if [ -f "$dir/SKILL.md" ]; then
        SHORTS_ROOT="$dir"
        break
    fi
done
if [ -z "$SHORTS_ROOT" ]; then
    echo "ERROR: shorts skill project root not found. Please run from the project directory or install with install.sh"
fi

Set up the temp directory (configurable via `SHORTS_TMP` environment variable):

SHORTS_TMP="${SHORTS_TMP:-/tmp/claude-shorts}"
mkdir -p "$SHORTS_TMP/clips"

10-Step Interactive Pipeline

Step 1: PREFLIGHT

Run safety checks on the input video:

bash "$SHORTS_ROOT/scripts/preflight.sh" INPUT_FILE [OUTPUT_DIR]

If preflight fails, report errors and stop. If warnings exist, report them and ask the user whether to proceed.

Also detect GPU capabilities:

bash "$SHORTS_ROOT/scripts/detect_gpu.sh"

Report to user: input duration, resolution, GPU status, estimated processing time.

Step 2: TRANSCRIBE

Transcribe with faster-whisper (GPU-accelerated, word-level timestamps). Audio extraction is handled internally by transcribe.py:

VENV="$HOME/.video-skill"
[ -d "$VENV" ] || VENV="$HOME/.shorts-skill"
source "$VENV/bin/activate"

python3 "$SHORTS_ROOT/scripts/transcribe.py" INPUT_FILE \
    --output $SHORTS_TMP/transcript.json

Output is dual-format JSON:

• `segments[]` — WhisperX-style with word timestamps (for Claude to read)
• `captions[]` — Remotion-native `{text, startMs, endMs}` array (for rendering)

Report to user: transcription time, word count, language detected.

Step 3: DETECT CONTENT TYPE

Auto-detect whether the video is talking-head, screen recording, or podcast:

python3 "$SHORTS_ROOT/scripts/detect_content.py" INPUT_FILE \
    --output $SHORTS_TMP/content_type.json

Report detected type to user. Ask if they want to override.

• **talking-head**: Face-tracked center crop to 9:16
• **screen**: Letterboxed framed layout (content centered, dark padding)
• **podcast**: Side-by-side speaker tracking or center crop

Step 4: ANALYZE — Claude Reads Transcript

Read the full transcript directly:

Read $SHORTS_TMP/transcript.json

Also load the scoring rubric:

Read $SHORTS_ROOT/references/scoring-rubric.md

Score 8-12 candidate segments (15-55 seconds each) on 5 dimensions:

| Dimension | Weight | What to look for | |-----------|--------|------------------| | Hook strength | 0.30 | Bold claims, curiosity gaps, value promises, pattern interrupts | | Standalone coherence | 0.25 | Makes complete sense without any context from the rest of the video | | Emotional intensity | 0.20 | Strong opinions, surprise reveals, humor, passion | | Value density | 0.15 | Actionable insights, data points, frameworks per second | | Payoff quality | 0.10 | Satisfying conclusion — punchline, reveal, call-to-action |

**Weighted score** = sum of (dimension_score * weight), scale 0-100.

For each candidate, identify:

• Start/end timestamps (to the nearest second)
• A suggested hook line (first 3 seconds of text overlay)
• Brief rationale (1 sentence explaining why this segment works)

**Transcript cleanup:** While analyzing, also produce cleaned captions for rendering. Read the `captions[]` array from transcript.json, then: 1. Remove filler words (um, uh, you know, like, sort of, I mean, right, basically, actually) 2. Fix obvious transcription errors based on surrounding context 3. Consolidate incomplete sentence fragments where appropriate 4. **Keep all timestamps unchanged** — only modify the `text` field

Write the cleaned transcript to `$SHORTS_TMP/transcript_cleaned.json` using the same JSON structure as transcript.json (both `segments` and `captions` arrays). The `captions` array should contain the cleaned text; copy `segments` as-is.

Step 5: PRESENT — Show Candidates Interactively

Present candidates in a formatted table:

| # | Time          | Dur  | Score | Hook                              | Why                                    |
|---|---------------|------|-------|-----------------------------------|----------------------------------------|
| 1 | 04:22 → 05:01 | 39s  | 87    | "Nobody talks about this..."     | Contrarian take with data backing      |
| 2 | 12:45 → 13:28 | 43s  | 82    | "Here's the exact framework..."  | Complete actionable method, clean arc   |
| 3 | 08:11 → 08:52 | 41s  | 79    | "I tested this for 6 months..."  | Personal story + surprising result     |

Then ask the user using AskUserQuestion: 1. **Which segments?** — "all", specific numbers, or "none, re-analyze" 2. **Caption style?** — bold (ALL CAPS pop-in), bounce (bouncy colorful), clean (minimal fade) 3. **Platform?** — youtube, tiktok, instagram, or all

Step 6: APPROVE — Interactive Adjustment Loop

After user selects segme