video-frame-analysis

Overview

Solidify "local video → ffprobe → timestamped frame extraction → contact sheet → OCR → summary.txt / ocr.json written to disk" into a reusable lightweight skill.

Suitable for agents like Codex, Claude Code, and OpenCode to handle local demo videos, screen recordings, and competitor page walkthroughs.

When to Use

Applicable to:

A user drops a local video and asks you to determine how far the product has gotten
Need to break a video into reviewable frame images, OCR text, and summary files
Need to do lightweight evidence extraction first, then manual judgment or further analysis
Want to avoid hand-assembling ffmpeg / tesseract commands every time

Not applicable to:

Online video downloading
Batch video queues
Real-time stream analysis
GUI or Web platforms
Heavyweight video understanding models

Quick Start

bash scripts/video-frame-analysis.sh /path/to/demo.mp4 ./out/demo-analysis

Default output:

ffprobe.txt
frame_*.jpg
contact.png
ocr.txt
ocr.json
summary.txt

Installation

Recommended to install directly into Codex / Claude Code / OpenCode:

npx skills add https://github.com/mason0510/video-frame-analysis-skills -g --all

After installation, restart the agent session, then trigger this skill as needed.

Input / Output

Input

Local video path
Output directory path

Output

Video metadata
Timestamped frame images
contact sheet
OCR text
OCR JSON
Summary text

Config

FRAME_INTERVAL_SECONDS=5 FRAME_WIDTH=420 OCR_LANG=eng \
  bash scripts/video-frame-analysis.sh ./demo.mp4 ./out/demo

Variable	Description	Default value
`FRAME_INTERVAL_SECONDS`	Frame extraction interval (seconds)	`8`
`FRAME_WIDTH`	Single frame width	`360`
`OCR_LANG`	OCR language candidates	`chi_sim+eng`

Project Files

scripts/video-frame-analysis.sh: main script
tests/test_video_frame_analysis.sh: minimal closed-loop test
references/output-format.md: output directory description
agents/openai.yaml: OpenAI/Codex side display information

Common Mistakes

Treat it as a heavyweight video understanding model
Extracting frames too densely causes OCR noise to spike
Ignore the contact sheet and only stare at the OCR text
Did not check whether the local machine has installed ffmpeg / tesseract