Video understanding for any model — native passthrough for small files, frame extraction + audio transcription fallback for large files. Use when the user asks to analyze, describe, or understand a video file (e.g. "what's in this video", "summarize this clip", "transcribe this recording").

Video Analysis Analyze video files using either native model understanding or frame extraction + transcription . ⚠️ URL input (YouTube / TikTok / IG / Apple Podcasts / Spotify / any link)? This skill takes a local file path . For a link, do NOT download the media. Route by input type, in order, and STOP at the first level that yields text: 1. Podcast episode link or episode name (Apple Podcasts, Spotify, Overcast, publisher site, or just "the a16z episode about X") → web search for the publisher's episode page / show notes / transcript page, then web crawler.scrape markdown(url) to read it. Most major shows publish text. 2. Video URL → web crawler.youtube video(url) / youtube transcript(url) (cloud captions, ~1 s, no download; TikTok has tiktok transcript ). 3. web fetch came back empty ( empty extraction , JS rendered page) → retry the same URL with web crawler.scrape markdown — it renders JS. Never conclude "no transcript" from an empty plain fetch. 4. Nothing found → report exactly that and ask. Local download + analyze video() runs ONLY when the user explicitly asks to download the media, or hands you a file. It is not a fallback: datacenter IPs hit bot checks and a 100 MB audio pull takes minutes and usually fails. Transcript text is speech only — never name a speaker from it alone (see web crawler's metadata first rule). How It Works Quick Start ⚠️ Invocation — do NOT use dotted imports. The directory name contains a hyphen ( video analysis ), so from skills.video analysis.exports import ... is a Python syntax error ( is parsed as minus). This is true for every hyphenated skill, not just this one. Use one of the two patterns below. Pattern A — from workspace root (recommended for scripts): Note: pass the video path workspace relative (analyze.py resolves it against WORKSPACE DIR ), even though you cd into the skill dir. Pattern B — inside a starchild clawd script: ❌ Do NOT exec(open('skills/video analysis/analyze.py').read()) — analyze.py uses file at import time, which is undefined under exec , so it crashes. Load it by file path with importlib.util.spec from file location if you must avoid both patterns above. Using the Exports Native Mode (small videos) For videos under the size threshold, the skill sends the full video to a model that supports native video input. The model sees every frame and hears the audio. Default model: google/gemini 3.1 flash lite — best price/quality for video. Model benchmark (6MB clip, vs gemini 3.1 pro preview baseline): Model Tier Cost Time Accuracy Notes google/gemini 3.1 flash lite budget ~$0.0014 8.1s ~88% ⭐ Default — cheapest + fastest google/gemini 3.5 flash std ~$0.0152 11.8s ~85% More detail, higher cost qwen/qwen3.6 plus budget ~$0.0058 44.2s ~95% Accurate but slow qwen/qwen3.6 flash budget ~$0.0027 16.6s ~80% Misreads subjects sometimes google/gemini 3.1 pro preview std ~$0.0199 19.7s 100% Baseline (best, most expensive) flash lite identifies the full scene, action sequence, and transitions correctly at ~14x lower cost than the Pro baseline. For maximum accuracy (exact character names, fine detail), switch default model to gemini 3.1 pro preview or gemini 3.5 flash in config/video analysis.yaml . Extraction Mode (large videos) For videos over the size threshold, the skill extracts keyframes and transcribes audio: Short videos (≤60s): One frame every N seconds (default: 2s) Long videos ( 60s): Scene change detection picks visually distinct frames Audio: Extracted and sent to Whisper for transcription Max frames: Capped at 30 (configurable) to control cost The agent receives frame image paths and transcript text, then feeds them to the current chat model as image attachments + context text. Configuration Edit config/video analysis.yaml (in the workspace) to customize. This file is created automatically on first use, only needs the keys you want to override, and survives skill updates . Do NOT edit skills/video analysis/config.yaml — that's the factory default and is overwritten on every skill auto update. The user file overlays it. Both the standalone skill and the chat "send a video" flow read this same config, so one edit changes the model everywhere. Available keys: Available Video Models Model Alias Tier Notes google/gemini 3.1 flash lite flash31 budget ⭐ Default, best price/quality google/gemini 3.5 flash gemini35 standard More detail, higher cost google/gemini 3.1 flash lite flash31 budget Cheapest option google/gemini 3.1 pro preview gemini standard Highest quality qwen/qwen3.6 flash qwenf budget Good alternative qwen/qwen3.6 plus qwen budget — minimax/minimax m3 mm3 standard — meta llama/llama 4 maverick maverick standard — meta llama/llama 4 scout scout budget — xiaomi/mimo v2.5 mimo standard — z ai/glm 5v turbo glm5v standard — minimax/minimax m2.7 mm27 budget Audio only, no image Agent Behavior When the user provides a video file (via upload or file path) and the current chat model does NOT support video: 1. Call analyze video(path, question) . 2. If result mode is "native" → return result["analysis"] directly. 3. If result mode is "extraction" → use result["frame paths"] as image references and result["transcript"] as context, then ask the current model to analyze based on the frames + transcript. When the current model DOES support video, the backend handles it natively via Phase 1 (base64 content block injection) — no need for this skill. Troubleshooting Problem Fix "File not found" Check path is workspace relative (e.g. output/videos/x.mp4 ) Native mode returns error Check default model in config/video analysis.yaml is valid No audio transcription Video may have no audio track; check has audio in result Too few frames extracted Lower scene threshold in config/video analysis.yaml (e.g. 0.15) Too many frames / high cost Reduce max frames or raise scene threshold

video-analysis