gif-full-view

Use when a GIF arrives — extract all frames for full motion.

SKILL.md

GitHub

🐦‍⬛ GIF Full View — See the whole animation, not frame 1

Trigger

Any GIF arrives in chat (Discord attachment, URL, meme, reaction). Default vision only shows the FIRST frame — the punchline lives in the motion. Run this and actually watch the GIF.

The Problem It Solves

  • A GIF freezes to frame 1 for the vision system — a "suspicious tea" GIF looks like a random close-up of a man making a peace sign.
  • The joke is the sequence: hand drops, expression shifts, stare intensifies.
  • Learned 8/11/26: trust the human's interpretation over the vision model's confident misidentification (it called a Dexter "I'm watching you" scene "Tyrese in 2 Fast 2 Furious" — the human knew better).

How To Do It

  1. Extract frames + build contact sheet (script exists):

    python3 /home/lumi/.hermes/profiles/vesper/scripts/gif2sheet.py <path-or-url> [--cols 4] [--max-frames 24]
    
    • Prints the sheet path on stdout (e.g. cache/gif-frames/sheet.png)
    • Saves individual frames under cache/gif-frames/frames/frame_NNN.png
    • Supports URLs (auto-downloads with browser UA) and local files
    • Defaults: 4 columns, max 24 frames sampled evenly
  2. Look at the sheet — vision_analyze on the sheet.png with a question like "Describe what happens across the frames — what motion or expression changes from frame 1 to the last frame?"

  3. Report the arc, not just the subject: what the hand does, how the expression shifts, the actual punchline.

Pitfalls

  • GIFs with MANY frames: cap at 24 frames for token sanity (script does this).
  • Tiny GIFs: 165x149 source gets upscaled fine in a contact sheet — don't worry.
  • Vision models hallucinate identities: it labeled a Dexter scene as 2 Fast 2 Furious. Report what you SEE (gesture, expression, motion), and if the user corrects the source, accept it — they know the scene.
  • First-frame-only trap: never judge a GIF by frame 1. The motion IS the joke.

Related

  • gif-search — finding GIFs via Tenor API (different job: search vs view).