Scenelens

/scenelens — Claude watches a video, smarter

You don't have a video input; this skill gives you one. Compared to a fixed-fps frame grab, scenelens:

1. Picks frames at scene changes — content-aware sampling instead of time-uniform sampling. Same frame budget, far better signal.
2. Runs OCR on every frame — on-screen text (slides, code, terminals, dashboards) is extracted as text alongside the image, so you don't burn vision tokens reading static pixels.
3. Auto-chunks long audio — Whisper's 25 MB cap no longer fails outright on long videos.

A Python script does all of this and prints a markdown report. You then Read each frame path to see the images and combine them with OCR + transcript to answer the user.

Step 0 — Setup preflight (silent on success)

Python interpreter: every python3 ... command in this skill is for macOS/Linux. On Windows, substitute python — python3 on Windows is the Microsoft Store stub and won't run the script.

Before every /scenelens call, verify dependencies and an API key are in place:

python3 "${CLAUDE_SKILL_DIR}/scripts/setup.py" --check

This is a [question]`.

How to invoke

Step 1 — parse user input. Separate the video source from any question. /scenelens https://youtu.be/abc what hook did they open with? → source = https://youtu.be/abc, question = what hook did they open with?.

Step 2 — run the script. Pass the source verbatim:

python3 "${CLAUDE_SKILL_DIR}/scripts/scenelens.py" ""

Optional flags:

• --mode auto|scene|fixed — frame selection strategy. Default auto: scene-aware first, fixed-fps fallback if scene changes are sparse. Force fixed for content with no hard cuts (e.g. a single-take talking head).
• --scene-threshold F — sensitivity (0-1, default 0.30). Lower = more frames captured. Bump to 0.20 for subtle visual changes.
• --start T / --end T — focus on a section (SS, MM:SS, HH:MM:SS).
• --max-frames N — lower the cap for tighter token budget.
• --resolution W — frame width in px (default 512; bump to 1024 only when the user must read tiny on-screen text and OCR isn't catching it).
• --no-ocr — skip the OCR pass. Use for content with no on-screen text (podcasts, interviews) to save a few hundred ms.
• --ocr-lang CODE — Tesseract language (default eng).
• --fps F — only applies in fixed-fps mode. Capped at 2 fps.
• --whisper groq|openai — force a specific backend. Default: prefer Groq when both keys exist.
• --no-whisper — disable Whisper entirely; frames-only if no captions.
• --sub-langs L1,L2 — caption languages in priority order (default en,en-US,en-GB,en-orig).
• --out-dir DIR — keep working files somewhere specific.

Step 3 — Read every frame path the script lists. The Read tool renders JPEGs directly as images. Read all frames in a single message (parallel tool calls). Each frame has a t=MM:SS timestamp. When OCR text is present, the report shows it inline — use that text directly instead of trying to read pixels.

Step 4 — answer the user. You now have THREE streams of evidence:

• Frames — what's on screen (chosen at scene cuts when possible)
• OCR — on-screen text, already extracted
• Transcript — what was said, with timestamps

If the user asked something specific, answer with timestamp citations. Otherwise summarize: structure, key visuals, what was said.

Step 5 — clean up. The script prints a working directory at the end. If the user isn't asking follow-ups, delete it with rm -rf .

Frame selection — why scene-aware matters

A 10-minute video with one demo and nine minutes of talking head:

• Fixed fps: 80 frames evenly spaced — 8 of them on the demo, 72 on the head.
• Scene-aware: dense around scene cuts — the demo frames cluster on UI changes, the head frames spread sparsely.

Same token cost, dramatically better signal. The default mode is auto: scene detection first, with automatic fallback to fixed-fps when fewer than 8 scene changes are detected (single-take videos, screen recordings of static UI). Use --mode fixed to force the legacy behavior; use --mode scene to disable the fallback.

Focusing on a section

When the user names a moment ("around 2:30", "the first 10 seconds", "the last 30 seconds"), pass --start / --end. Frame budget tightens around the range, transcript filters to the same window, frame timestamps stay absolute (real video timeline).

python3 "${CLAUDE_SKILL_DIR}/scripts/scenelens.py" video.mp4 --start 50 --end 60
python3 "${CLAUDE_SKILL_DIR}/scripts/scenelens.py" "$URL" --start 2:15 --end 2:45
python3 "${CLAUDE_SKILL_DIR}/scripts/scenelens.py" "$URL" --start 1:12:00

Transcription

1. Native captions (free, preferred). yt-dlp pulls manual or auto-generated subtitles when available.
2. Whisper API fallback. If captions are missing, the script extracts mono 16 kHz mp3 audio (~480 kB/min) and uploads it to Groq's whisper-large-v3 (preferred) or OpenAI's whisper-1.
3. Auto-chunking for long audio. Audio >24 MB is split into chunks under the 25 MB API cap, each transcribed separately, then merged with offset timestamps. A 4-hour podcast no longer fails — it just makes more API calls.

Both keys live in ~/.config/scenelens/.env. Unlike skills that fall back to a project-local .env, scenelens reads ONLY from ~/.config/scenelens/.env and process env — to avoid silently picking up keys from random project directories.

Failure modes

• Setup preflight failed → run python3 "${CLAUDE_SKILL_DIR}/scripts/setup.py" (auto-installs ffmpeg/yt-dlp via brew on macOS, scaffolds .env). For an API key, ask the user via AskUserQuestion and write it to ~/.config/scenelens/.env.
• No transcript available → captions missing AND (no Whisper key OR Whisper API failed). Proceed frames-only and tell the user.
• --mode scene returned no frames → the video has no detectable scene changes. Re-run with --mode auto (default) or --mode fixed.
• OCR not available → tesseract not installed. The skill keeps working; the report says OCR: unavailable. Mention to the user that installing tesseract would unlock the OCR sidechannel.
• Whisper request fails → printed to stderr (likely: invalid key, rate limit, or all chunks failed). Retry with --whisper openai if Groq failed (or vice versa).

Token efficiency

Frames dominate token cost. Order of magnitude:

• 80 frames @ 512 px ≈ 50-80 k image tokens.
• Transcript ≈ a few thousand tokens for a 10-min video.
• OCR text is cheap — a few hundred to a couple thousand tokens.
• Bumping --resolution to 1024 roughly 4x's image tokens per frame. Only do it when OCR doesn't catch what's needed.

If you already watched a video this session and the user asks a follow-up, do not re-run the script — you already have the frames + OCR + transcript. Answer from context.

Security & Permissions

What this skill does:

• Runs yt-dlp locally to download the video and pull native captions (public data, request goes to whatever host the URL points at).
• Runs ffmpeg / ffprobe locally to extract frames and (when Whisper is needed) mono 16 kHz audio.
• Runs tesseract locally on the extracted JPEG frames (no network call).
• Sends the audio clip to Groq's Whisper API (api.groq.com/openai/v1/audio/transcriptions) when GROQ_API_KEY is set.
• Sends the audio clip to OpenAI's audio transcription API (api.openai.com/v1/audio/transcriptions) when OPENAI_API_KEY is set and Groq is not, or when --whisper openai is forced.
• Writes downloaded video, frames, audio, and intermediate transcripts to a working directory under the system tmp dir (or --out-dir) so Claude can Read them.
• Reads / creates ~/.config/scenelens/.env (mode 0600 on POSIX) for the Whisper API key(s) and a SETUP_COMPLETE marker.

What this skill does NOT do:

• Does NOT upload the video itself anywhere — only the extracted audio goes out, and only when captions are missing AND Whisper is enabled.
• Does NOT read .env from the current working directory (would silently pick up keys from random project dirs).
• Does NOT write to system paths — --out-dir is rejected if it resolves under /etc, /bin, /sbin, /usr/bin, /usr/sbin, or /boot.
• Does NOT log API keys to stdout, stderr, or output files.
• Does NOT access platform accounts (no login, no cookies, no posting).
• Does NOT share keys between providers (Groq key only goes to api.groq.com, OpenAI key only to api.openai.com).

Bundled scripts: scripts/scenelens.py (entry), scripts/download.py (yt-dlp wrapper), scripts/frames.py (scene + fixed-fps extraction), scripts/ocr.py (Tesseract wrapper), scripts/transcribe.py (VTT parser), scripts/whisper.py (Groq/OpenAI clients with chunking), scripts/setup.py (preflight + installer).

Review scripts before first use to verify behavior.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ravindranathpathi
• Source: ravindranathpathi/scenelens
• License: MIT
• Homepage: https://scenelens.tech

Install and usage instructions live in the source repository linked above.