Claude traffic-CV starter pack

8 prompts + a free open-source path to a demo (not for tickets)

Build a prototype that draws boxes on vehicles and estimates speed from a fixed camera. This is for learning and demos — not for issuing fines, enforcement, or selling to a city. Real speed enforcement needs calibrated gear, certification, and a lawyer. The viral $317K story in the video is not confirmed.

1What you need

28 Claude prompts (copy/paste, fill brackets)

  1. "I'm a beginner. Plan a minimal traffic counter: fixed webcam, Python, OpenCV, YOLOv8n. List folders, packages, and the first 5 files to create. No cloud API for inference."
  2. "Write a Python script that opens webcam [index 0], runs YOLOv8n, draws boxes only for classes car/truck/bus/motorcycle, shows FPS. Add clear comments."
  3. "Add a vertical tripwire line at x = [frame_width//2]. Count each track ID that crosses it once. Print counts per class."
  4. "Estimate speed: assume pixels-per-meter = [PPM] for my camera height. For each track, speed_kmh = distance_m / time_s * 3.6. Show the number above the box. Warn that this is uncalibrated."
  5. "Add a side panel: live counts, total, and an Export CSV button that writes timestamp, track_id, class, speed_kmh."
  6. "Rewrite my script to use a video file [path] instead of the webcam, same UI, so I can demo without standing on a balcony."
  7. "List 10 ways this demo is NOT a legal speed radar (calibration, angle, lens distortion, weather, plate OCR, chain of custody). Keep it blunt."
  8. "Give me a 60-second demo script I can narrate while the app runs: what to point at on screen, what NOT to claim."

3Calibration reality check (say this out loud in any demo)

4If someone asks about the $317K claim

Point them to: Times of India, 8 Jul 2026 — the sale and student identity "have not been independently verified" / "no confirmed public evidence."