Stats
Actions
Tags
From Docs from videos
Turn a product demo or screen-recording video into a structured feature inventory plus matching screenshots, using the Gemini API. Use this skill whenever the user has a video (demo, walkthrough, tutorial, screen capture) and wants to know what it shows — pull out the features, get a feature list, or grab annotated screenshots — even if they don't mention Gemini by name. Triggers on "analyze this demo video", "what features are in this recording", "extract screenshots from this walkthrough", or "turn this MP4 into a feature list". It captures features that are SPOKEN in the narration as well as ones only SHOWN on screen, and pulls a screenshot for each at the right moment. (To go further and write actual documentation pages from a video, use the video-to-docs skill, which builds on this one.)
How this skill is triggered — by the user, by Claude, or both
Slash command
/doc:extract-product-doc-from-videosThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Going through a demo video by hand to capture what it shows is slow: you scrub
Going through a demo video by hand to capture what it shows is slow: you scrub back and forth, transcribe the narration, note what's on screen, and screenshot each feature. This skill does that first pass automatically. Gemini watches the whole video — audio and visuals — and returns a structured list of features, each with a description and the timestamps where it's best shown. The script then grabs a screenshot at each of those moments. You get a Markdown draft, a JSON file, and a folder of screenshots.
It captures two kinds of features on purpose: things the narrator talks about
(even off-screen) and things only shown on screen. The source field on each
feature tells you which.
Nothing is installed globally. The script declares its Python dependencies
inline (PEP 723) and runs with uv, which builds an isolated, cached
environment on the fly. You need:
uv on PATH (uv --version). Single binary; see astral.sh/uv.GEMINI_API_KEY (or GOOGLE_API_KEY) in the environment.
Uploads via the File API, so it handles large videos.gcloud auth application-default login done and a
project set (gcloud config set project <id>, GOOGLE_CLOUD_PROJECT, or
--project). Vertex has no File API, so the video is sent inline —
keep clips under ~20 MB in this mode, or use API-key mode for big files.ffmpeg does not need a system install — the bundled imageio-ffmpeg
dependency provides a static binary (a system ffmpeg on PATH is used if one
exists).
uv run "${CLAUDE_PLUGIN_ROOT}/skills/extract-product-doc-from-videos/scripts/extract_features.py" <video> \
[--model gemini-3.1-pro-preview] [--format webp] [--gcs-bucket NAME] [--output <dir>] [--fps N] [--quiet]
<video> — path to the video file (mp4, mov, webm, …).--model — defaults to gemini-3.1-pro-preview (best feature coverage and
accuracy in testing, ~3× cheaper video tokenization than the 2.5 models). Use
gemini-3.5-flash for the fastest, cheapest pass.--format — png (default) or webp. Use webp for docs sites.--gcs-bucket — on Vertex AI, a GCS bucket to upload the full-quality video
to (no compression). Defaults to $GEMINI_VIDEO_BUCKET. Strongly recommended for
reading on-screen text accurately.--output — output folder. Defaults to <video-name>-features/.--fps — frames-per-second Gemini samples (default ~1). Raise it for
fast-moving UIs where 1 FPS misses things.The script prints the output directory path on stdout when done.
features.md — one section per feature: title, source (voice / visual /
both), description, and the embedded screenshot(s).features.json — the same data, structured.screenshots/ — one image per timestamp, named
<index>-<feature-slug>-<timestamp>.<png|webp>.--fps if one missed.--gcs-bucket (or $GEMINI_VIDEO_BUCKET) to
upload the original to Cloud Storage and analyze it at full quality — this gives
the model the sharpest view of on-screen text and detail. The uploaded object is
deleted after the run. Without a bucket, a video over ~18 MB falls back to a
downscaled inline copy for analysis only (tune with --max-upload-mb).
Screenshots are always grabbed from the original, so they keep full resolution
in every mode — never pre-downscale the input yourself.npx claudepluginhub kapptivate/ktm-docs-jamdesk --plugin docFetches up-to-date documentation from Context7 for libraries and frameworks like React, Next.js, Prisma. Use for setup questions, API references, and code examples.
Applies a firm's KYC/AML rules grid to parsed onboarding records: assigns risk rating, checks required documents, outputs rule outcomes with citations, and routes for escalation.
Generates daily or weekly digests of activity from connected sources (chat, email, docs, tasks, CRM), highlighting action items, decisions, mentions, and project updates.