AgentSkillsCN

Docs

文档

SKILL.md

SKILL.MD

Common task shortcuts for CV Rapid Response. These patterns save tokens by eliminating redundant exploration.

Quick File Access

Core files (use Read directly):

code
backend/videostreams.py       # Main processing script
backend/.env                  # Configuration
CLAUDE.md                     # Project docs
synchronized_security_report.json  # Latest detection output

Common Tasks

1. Run Video Processing

Command: cd backend && python3 videostreams.py

Before running:

  • •Verify VIDEO_PATH exists: ls backend/video/
  • •Check .env has ROBOFLOW_API_KEY
  • •No need to read entire file unless modifying

After running:

  • •Check output: synchronized_security_report.json in project root
  • •Annotated video in project root with prefix annotated_

2. Change Detection Target

File: backend/videostreams.py:23 Edit: TARGET_OBJECT = "person" → any COCO class (car, bicycle, dog, etc.) No restart needed - just re-run script

3. Add Video Source

File: backend/videostreams.py:22 Pattern: Add path to VIDEO_SOURCES list

python
VIDEO_SOURCES = [VIDEO_PATH, "/path/to/video2.mov"]

4. Check Dependencies

Installed: pip list | grep -E "(inference|supervision|opencv)" Missing: pip install inference supervision opencv-python Virtual env: .venv exists, may need activation

5. Analyze Detection Report

Location: synchronized_security_report.json Structure:

json
{"frame_N": {"cam_1": {"detected": "YES/NO", "center": [x, y]}}}

Quick stats: cat synchronized_security_report.json | grep -c '"detected": "YES"'

6. Model Configuration

Default: yolov8n-640 (COCO trained, 80 classes) Change model: Edit model_id in videostreams.py:67 Custom model: Set ROBOFLOW_MODEL in .env Variants: yolov8s-640 (more accurate), yolov8m-640 (balanced)

Token-Saving Patterns

Instead of exploring:

  1. •User asks about detection → Read videostreams.py:29-63 (synchronized_sink function)
  2. •User asks about output → Read synchronized_security_report.json
  3. •User asks about config → Read backend/.env and videostreams.py:16-24

Parallel operations when uncertain:

python
# Good: Check multiple locations at once
Glob("**/*.env")
Glob("**/config*.py")
Read("backend/videostreams.py")

Avoid redundant reads:

  • •Cache file structure mentally from first read
  • •Jump to line ranges for specific changes
  • •Use Edit tool for known modifications

Environment Setup

Python version: 3.10 (see .venv) Working directory: Project root /Users/jago/Downloads/Auto-GPT/cv_rapid_response Backend scripts: cd backend before running Output location: Project root (not backend/)

Debugging Shortcuts

No video file:

  • •Check: ls backend/video/IMG_3501.MOV
  • •Fix: Add video to backend/video/

API key error:

  • •Check: grep ROBOFLOW_API_KEY backend/.env
  • •Fix: Add ROBOFLOW_API_KEY=your_key to backend/.env

Import errors:

  • •Check: .venv/bin/python3 -c "import inference"
  • •Fix: cd backend && ../.venv/bin/pip install inference

Empty detections:

  • •Verify TARGET_OBJECT is valid COCO class
  • •Check video has target object
  • •Try different model (yolov8s-640 for better accuracy)