video-feed-ocr

Video Feed OCR

If the task is extracting on-screen text from video, use this skill first.

Overview

Run frame-by-frame OCR for general video analysis using one self-contained Python script in this skill.

The script supports two OCR engines:

• Fast path: Tesseract via --fast (lower latency, may be less accurate)
• High quality: Local multimodal LLM (higher quality on complex layouts)

If fast output is noisy, missing lines, or misread, rerun without --fast to use the LLM path.

Workflow

1. Confirm the input video exists and is readable.
2. Run python3 scripts/video_feed_ocr.py with an output directory under /tmp or the current workspace.
3. For quick/cheap pass, add --fast.
4. Review ocr_raw.txt for full extracted text.
5. Optionally filter with --pattern to generate ocr_hits.txt.
6. Report both what was detected and confidence caveats (OCR typos, low contrast, motion blur, model hallucination risk).

Quick Start

Fast (Tesseract):

python3 scripts/video_feed_ocr.py \
  --input /path/to/video.mp4 \
  --output-dir /tmp/video-ocr \
  --fast

High Quality (LLM):

python3 scripts/video_feed_ocr.py \
  --input /path/to/video.mp4 \
  --output-dir /tmp/video-ocr

Execution Context

Always run this skill in escalation permission mode.

Common Options

• --fps N: frames per second to sample (default 1)
• --interval S: sample one frame every S seconds (alternative to --fps)
• --frame-limit N: stop after N frames (for quick previews)
• --pattern REGEX: write filtered matches to ocr_hits.txt
• --fast: use Tesseract instead of the LLM path
• --lang LANG: Tesseract language in fast mode (default eng)
• --psm N: Tesseract page segmentation mode in fast mode (default 6)
• --task-prompt TEXT: optional custom OCR extraction instruction for LLM mode
• --timeout SEC: timeout per OCR call

Model is hardcoded in the script to qwen/qwen3-vl-4b for LLM mode.

Default OCR prompt used when --task-prompt is omitted:

Extract all text exactly as seen.
Keep reading order.
Keep line breaks.
Do not paraphrase.
Do not translate.
If unsure add (?).
If unreadable write [illegible].

Use:
Paragraphs for normal text.
Bullets for obvious lists.
Tables only if clearly a table.

Return Markdown only.

Examples:

# Every 2 seconds, faster sampling
python3 scripts/video_feed_ocr.py --input input.mp4 --output-dir /tmp/ocr --interval 2

# Fast OCR with explicit Tesseract options
python3 scripts/video_feed_ocr.py \
  --input input.mp4 \
  --output-dir /tmp/ocr \
  --fast \
  --lang eng \
  --psm 6

# If fast output is poor, rerun with LLM + optional custom prompt
python3 scripts/video_feed_ocr.py \
  --input input.mp4 \
  --output-dir /tmp/ocr \
  --task-prompt "Extract subtitles only. Preserve line breaks."

# Look for emails, URLs, ticket ids, or any custom regex
python3 scripts/video_feed_ocr.py \
  --input input.mp4 \
  --output-dir /tmp/ocr \
  --pattern 'https?://[^ ]+|[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+'

Resource

• scripts/video_feed_ocr.py