agentsclimarketplace
Sign in

Glideit

Skill Imhari14/glideit/skills/glideit

Watch a long video (URL or local path) and answer questions about it. Downloads with yt-dlp, builds a storyboard montage + transcript of the WHOLE video, then zooms into the relevant window with dense high-res frames and OCR of on-screen code. Use for lectures, tutorials, screencasts, conference talks, screen recordings — anything more than a couple of minutes. No external model is called: YOU read the frames.From its SKILL.md

Install
npx -y skills add Imhari14/glideit --skill glideit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

4.9 KB, ~1.2k tokens by cl100k_base, as published. Nobody here has run it

glideit — watch a long video

Give yourself eyes for a full-length video without drowning your context. Two phases: map the whole video cheaply, then zoom into the part that matters. Never try to Read every frame of a long video — map first, then zoom.

$CLAUDE_PLUGIN_ROOT is this plugin's directory. Use it so the script resolves no matter the working directory. (Running standalone from the repo? Use scripts/glideit.py instead.)

Workflow

  1. Map the whole video.

    python "$CLAUDE_PLUGIN_ROOT/scripts/glideit.py" "<url-or-path>"

    Prints a transcript path (transcript.txt) and one or more storyboard montages — grids of timestamp-labelled thumbnails spanning the whole runtime. Read the transcript, then Read the storyboards — all in ONE message (parallel Reads), never one per turn. The frames are what ground your answer in what's actually on screen; the transcript alone can't do that. One shortcut: if the user already names a moment ("what happens at 12:30?"), skim the transcript around it and zoom straight there.

  2. Decide where to look. From the transcript + storyboards, pick the window that answers the user's question (e.g. the code appears around 12:00–13:30).

  3. Zoom in.

    python "$CLAUDE_PLUGIN_ROOT/scripts/glideit.py" "<url-or-path>" --start 12:00 --end 13:30 --resolution 1024

    Prints dense, high-res frame paths for that window plus OCR of any on-screen code (in zoom_manifest.json). Read each frame. For code, trust the OCR text for exact characters and the frame for layout.

  4. Answer grounded in the frames and transcript you actually read.

Flags

  • --detail fast|balanced|deep — map density (deep also OCRs each map frame).
  • --start / --end — zoom window (SS, MM:SS, or HH:MM:SS).
  • --timestamps 3:12,3:40 — exact moments instead of a window.
  • --fps F — zoom sampling rate (frames/sec), e.g. 2 = every 0.5s. Overrides the default ~1 frame/3s to catch fast motion (demos, UI, gestures, a flashing error) in a short window — cheap because the window is already small.
  • --resolution N — bump to 1024+ for legible on-screen code.
  • --budget N — storyboard thumbnail count (default 64).
  • --no-ocr — skip the OCR sidecar.
  • --cards — emit cards.json (per-card text + narration + timing) and a hyperframes_scaffold.html starter, for recreating/remixing the video.
  • --note "..." — append a note to this video's persistent notes.md, then exit.
  • --refresh — ignore the cache and rebuild the map/zoom.

Reuse — don't re-scan, don't re-read

The map and every zoom are cached under .glideit/<hash>/. Re-running the same map or zoom returns instantly ((cached ...)) with no re-extraction — pass --refresh only when you truly want to rebuild. To keep tokens low:

  • Lean on the transcript (cheap text) first; Read frames only for visual detail the transcript can't provide.
  • Don't re-Read a storyboard or frame already in your context this session.
  • After you understand the video, save a digest with python "$CLAUDE_PLUGIN_ROOT/scripts/glideit.py" "<url-or-path>" --note "..." (one --note per point; appends to <workdir>/notes.md). Every later map/zoom prints prior notes: … — Read that text first and only re-read an image when you need a visual detail the notes don't already capture.

Recreate or remix a video (with HyperFrames)

To rebuild a video as an editable template, run with --cards. It writes:

  • cards.json — one entry per detected card: timestamp, t_start/t_end, on-screen text (OCR), the narration over it, and a reference frame.
  • hyperframes_scaffold.html — a starter HyperFrames composition (one timed clip per card) to flesh out with the hyperframes skills, then npx hyperframes render.

Read cards.json + the reference frames, apply the user's changes (new content, brand, language, layout), refine the scaffold, render, then run glideit on the rendered MP4 to review your own output and iterate.

Notes

  • Everything runs locally (ffmpeg / yt-dlp / tesseract). No model API is called.
  • Captions are used when present; install faster-whisper for caption-less/local videos. OCR needs tesseract on PATH (degrades gracefully if absent).
  • The download, map, and zooms are all cached under .glideit/<hash>/.
  • Run python "$CLAUDE_PLUGIN_ROOT/scripts/setup.py" once to verify local tools.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Gives 0 of the 12 instructions most video audio skills give in ~1.2k tokens

Counted across 619 of the 725 authors here whose files we hold, read 2026-09-06

  • Read product marketing context firstin 13 of 619, across 7 files
  • Define the core visual thesis in one sentencein 11 of 619, across 3 files
  • Break the concept into 3 to 6 scenesin 11 of 619, across 3 files
  • Render the smallest working version firstin 11 of 619, across 3 files
  • Start with a low-quality smoke test renderin 11 of 619, across 3 files
  • Add captions for accessibility and engagementin 11 of 619, across 5 files
  • Write the scene outline before writing codein 11 of 619, across 3 files
  • Specify subject, action, camera, style, and moodin 11 of 619, across 5 files
  • Decide what each scene provesin 10 of 619, across 2 files
  • Export one clean thumbnail framein 10 of 619, across 2 files
  • Pick the right tool for the jobin 10 of 619, across 4 files
  • Run the test suite before proposing a fixin 8 of 619, across 7 files

Said here and by no other author read

  • map the whole video first
  • read the transcript and storyboards in parallel
  • decide where to look using transcript and storyboards
  • zoom into the relevant video window
  • trust OCR text for exact characters
  • lean on the transcript first

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.