nano-banana

Personal collection of Claude Code skills for PR automation and creative tasks.

Installation

Step 1: Add the Marketplace

Add this marketplace to Claude Code:

/plugin marketplace add devonjones/devon-claude-skills

Step 2: Install Plugins

Install the plugins you want to use:

# Install PR review loop
/plugin install pr-review-loop@devon-claude-skills

# Install image generation
/plugin install nano-banana@devon-claude-skills

# Install YouTube transcripts
/plugin install youtube-transcript@devon-claude-skills

# Install YouTube screenshotter
/plugin install youtube-screenshotter@devon-claude-skills

# Install YouTube synthesizer (recommended: install all three YouTube plugins together)
/plugin install youtube-synthesizer@devon-claude-skills

# Or install all
/plugin install pr-review-loop@devon-claude-skills nano-banana@devon-claude-skills youtube-transcript@devon-claude-skills youtube-screenshotter@devon-claude-skills youtube-synthesizer@devon-claude-skills

Step 3: Verify Installation

Check that the plugins are installed:

/plugin list

The skills will be automatically available for Claude to use.

Available Plugins

pr-review-loop

Automate PR review feedback loops with Gemini Code Assist, Cursor Bugbot, and Claude fallback.

Install:

/plugin install pr-review-loop@devon-claude-skills

Features:

• •Autonomous review loop execution as Task agent
• •Multi-bot support (Gemini Code Assist, Cursor Bugbot, Claude)
• •Automatic comment resolution with reply templates
• •Rate limit detection and fallback
• •Skeptical review evaluation (not all suggestions should be implemented)
• •Diminishing returns detection with "ONE MORE LOOP" verification

Usage:

# Spawn autonomous review agent
Task tool:
  subagent_type: general-purpose
  model: sonnet
  description: PR review loop for #<PR>
  prompt: Execute the PR review feedback loop for PR #<PR>...

nano-banana

Generate, edit, and compose images using Google's Gemini 3 Pro Image model.

Install:

/plugin install nano-banana@devon-claude-skills

Features:

• •Text-to-image generation with aspect ratio control
• •Image editing with natural language instructions
• •Multi-image composition (up to 14 reference images)
• •Real-time data visualization with Google Search grounding
• •Support for 1K, 2K, and 4K resolutions

Usage:

# Generate an image
scripts/generate_image.py "a serene landscape" output.png --aspect 16:9

# Edit an existing image
scripts/edit_image.py input.png "add a sunset" output.png

# Compose multiple images
scripts/compose_images.py "combine these" output.png img1.png img2.png img3.png

youtube-transcript

Extract transcripts (auto-generated or uploaded) from YouTube videos.

Install:

/plugin install youtube-transcript@devon-claude-skills

Features:

• •URL parsing for watch?v=, youtu.be/, youtube.com/shorts/, youtube.com/embed/, and bare 11-character video IDs
• •Plain text or [MM:SS] / [HH:MM:SS] timestamped output
• •Language preference via --lang
• •Direct file output via -o PATH
• •Dependencies installed in an ephemeral per-invocation venv via uv run --script (no persistent Python install pollution)

Usage:

# Plain transcript to stdout
scripts/get_transcript.py "https://www.youtube.com/watch?v=VIDEO_ID"

# With timestamps
scripts/get_transcript.py "VIDEO_ID" --timestamps

# Specific language, write to a file
scripts/get_transcript.py "VIDEO_ID" --lang en -o transcript.txt

youtube-screenshotter

Download a YouTube video at 720p and extract frames at specified timestamps with perceptual hashes. Mechanical only — pairs with youtube-synthesizer (which adds the smart bisection + visual-diff judgment) or any other downstream tool that needs frames + pHashes.

Install:

/plugin install youtube-screenshotter@devon-claude-skills

Features:

• •Single extract.py CLI batches yt-dlp download + ffmpeg frame extraction + pHash computation
• •Manifest JSON output: video metadata block + per-frame entries {timestamp, frame_path, phash}
• •Cached on disk: video file by <video_id>.mp4, frames by t<ms>.png — repeat invocations are cheap
• •Sub-scripts also runnable independently: video.py (download/metadata), frames.py (frame extraction), phash.py (compute / 3-band Hamming classifier)
• •Dependencies via uv run --script (yt-dlp, imagehash, Pillow)

Usage:

# Extract frames at given timestamps + emit manifest JSON
scripts/extract.py "https://www.youtube.com/watch?v=VIDEO_ID" -t 1 -t 30 -t 500 -o ./out

# Just metadata, no download
scripts/video.py "VIDEO_ID" --no-download

# Compute pHash of one image; compare two
scripts/phash.py compute frame.png
scripts/phash.py compare frame_a.png frame_b.png

youtube-synthesizer

Ingest a YouTube video as a faithful-capture literature note in an Obsidian vault. Combines transcript + canonical visual scenes + structured extraction (TL;DR, takeaways, references, code, tags). The smart layer on top of youtube-transcript and youtube-screenshotter — drives a bisection + visual-diff judgment loop using the agent's native vision capability (no separate API calls).

Install:

/plugin install youtube-synthesizer@devon-claude-skills

Recommended companions: youtube-transcript and youtube-screenshotter (this skill orchestrates them).

Features:

• •Faithful capture, no editorializing — the entry preserves what the video communicated; weaving content into your mental model stays in your hands
• •Per-note folder under <vault>/sources/videos/<ingested-date>-<sanitized-title>/ so each entry self-contains its assets and is portable
• •Frontmatter follows a stable base schema shared with future article/pdf/substack source-skills, plus YouTube-specific extras (channel_id, playlist_id, chapter_markers, etc.) — designed for a future sqlite3 corpus indexer
• •Hierarchical (chapter → scenes) when chapter markers exist; flat for short videos
• •Skip-by-default if entry already exists; --rerun to overwrite
• •All inference runs on your existing Claude Code session — no separate ANTHROPIC_API_KEY

Usage:

The skill is invoked via natural language or slash command, not a CLI script. Provide a video URL and a target vault path:

Use the youtube-synthesizer skill to ingest
https://www.youtube.com/watch?v=VIDEO_ID into ~/Obsidian/woodworking

Optional flags interpreted by the skill: --rerun to overwrite an existing entry, --why "<reason>" to seed the why_ingested frontmatter field.

Migration Notice

These skills were previously hosted in separate repositories:

• •`skill-pr-review-loop` → Archived, redirects here
• •`skill-nano-banana` → Archived, redirects here

Prerequisites

pr-review-loop

• •gh CLI authenticated with GitHub
• •git installed
• •pre-commit (optional, auto-detected)

nano-banana

• •Python 3.8+
• •GEMINI_API_KEY environment variable
• •Dependencies: pip install google-genai Pillow pyyaml

youtube-transcript

• •uv — installs the Python deps in an ephemeral venv per invocation. Install with curl -LsSf https://astral.sh/uv/install.sh | sh or pip install uv.
• •Network access to youtube.com.

youtube-screenshotter

• •uv — installs Python deps (yt-dlp, imagehash, Pillow) in an ephemeral venv per invocation.
• •ffmpeg — required for frame extraction and yt-dlp's video+audio merge. sudo apt install ffmpeg (Linux), brew install ffmpeg (macOS).
• •Network access to youtube.com and googlevideo.com for downloads.

youtube-synthesizer

• •youtube-transcript and youtube-screenshotter plugins installed (transitively requires uv and ffmpeg)
• •Network access to YouTube
• •A target Obsidian vault path (writable). Per-note folder convention; the skill writes only under <vault>/sources/videos/.

License

MIT - See LICENSE for details.