Video Loop

Video Loop

Create a seamless looping background video from any source clip. A crossfade blends the last and first frames so the `` transition is invisible, then the pipeline encodes web-optimized MP4 + WebM.

All ffmpeg work happens in scripts/video-loop.sh — this skill validates inputs, optionally helps pick a good fade duration, invokes the script, and turns the script's summary into a human-readable report.

Parameters

| Flag | Default | Description | |------|---------|-------------| | -d | 1.5 | Crossfade duration in seconds | | -q | 26 | H.264 CRF — 18 = high quality, 28 = smaller file | | -w | 32 | VP9 WebM CRF | | -o | Same as input | Output directory | | -p | off | Also extract a poster frame (first frame as JPEG) | | -C | off | Disable the crossfade — optimize and encode only (-n accepted as a deprecated alias) |

> No -s/-S save-mode flag. Unlike media-cluster siblings (/audio-loop, /markitdown), video-loop outputs are web assets (MP4 / WebM / poster) the user drops directly into their site's video directory. No downstream skill consumes video files via -f, so the ~/.agents/output/// convention would add ceremony without enabling anything — -o is the direct path.

Workflow

1. Validate tools

Confirm ffmpeg and ffprobe are available: command -v ffmpeg ffprobe. If either is missing, stop and ask the user to install (macOS: ! brew install ffmpeg, Debian/Ubuntu: ! sudo apt install ffmpeg). Never auto-install.

2. Analyze loop quality (optional but recommended)

Before committing to a crossfade, extract the first and last frames and compare them — this is the one step in the pipeline that genuinely benefits from the model's visual reasoning:

ffmpeg -y -ss 0 -i  -frames:v 1 -update 1 /tmp/vl-first.png
ffmpeg -y -sseof -0.1 -i  -frames:v 1 -update 1 /tmp/vl-last.png

Read both frames. If they're nearly identical, suggest -C (the source already loops cleanly, crossfade would only blur motion). If they differ significantly, proceed with the default fade or recommend a longer -d for more gradual blending.

3. Run the pipeline

$SKILL_DIR = this skill's folder — ${CLAUDE_SKILL_DIR} in Claude Code, the directory containing this SKILL.md elsewhere.

bash "$SKILL_DIR"/scripts/video-loop.sh  [flags]

The script does the rest: probes metadata, builds the lossless loop intermediate when crossfade is requested, encodes MP4 (H.264 + faststart) and WebM (VP9), optionally extracts a poster frame. It emits a machine-readable summary on stdout, one RESULT: key=value line per fact.

4. Report

Parse the RESULT: lines to compose:

• A size table comparing input vs outputs (absolute MB and delta percent)
• A ready-to-paste HTML snippet for the user's site

Report template:

| File | Size | Codec | vs Original |
|------|------|-------|-------------|
| `*-opt.mp4` | X MB | H.264 | -XX% |
| `*.webm` | X MB | VP9 | -XX% |
| `*-poster.jpg` | X KB | JPEG | — |

HTML snippet:

WebM first — browsers pick the first supported source.

Why crossfade-first matters

Useful background for debugging "I still see a jump" reports. The script places the xfade at the start of the output, not the end:

• [xfade] (F seconds): blends A[end] → A[start]. First frame = A[D-F], last frame = A[F]
• [middle] (D−2F seconds): untouched frames from A[F] to A[D-F]

At the loop boundary, [middle] ends at A[D-F] and the player loops back to [xfade], which starts at the same frame — invisible transition. If the xfade were placed at the end ([middle][xfade]), the xfade would end at A[F] but loop back to A[0] — a visible jump of F seconds.

If the user reports a visible jump despite running the skill, first check whether they pointed at an already-encoded file (re-running re-encodes but doesn't re-loop). If the source is correct, try a longer -d — short fades are more visible on high-motion content.

Rules

• NEVER upscale — the script encodes at source resolution
• NEVER re-encode an already-encoded output (double compression). Always start from the original source
• Audio is always stripped (-an) — background videos are muted
• The script rejects -d >= duration/2 — a too-long fade would leave the middle segment empty
• Always report original vs output sizes so the user sees the quality/size tradeoff

Source & license

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

• Author: coroboros
• Source: coroboros/agent-skills
• License: MIT
• Homepage: https://coroboros.com

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