feat: add supervision AI coding skills for Claude Code and Cursor

[greeshmab21]

Summary

Adds an official AI coding skill for supervision, so agents like Claude Code and Cursor can use the library correctly instead of guessing at API names. Closes #2244.

The skill is organized as a top-level SKILL.md plus focused reference files under skills/references/, each targeting a spot where agents commonly hallucinate methods/params or pick the wrong tool.

What's included

• skills/SKILL.md — entry point. The standard detect + annotate pattern (model → sv.Detections → annotate → save), plus a decision guide for sv.process_video vs. InferencePipeline: file-in/file-out with a simple callback vs. live/streaming sources with threaded inference. Links out to each reference file.

• skills/references/detection.md — building sv.Detections from ultralytics/inference/SAM/transformers, key attributes (xyxy, confidence, class_id, tracker_id, mask), and filtering via boolean-mask indexing (detections[detections.class_id == 0]) instead of the nonexistent .filter().

• skills/references/annotators.md — the full annotator class list with correct parameter names, the chaining/compose pattern for layering multiple annotators, and the BoundingBoxAnnotator → BoxAnnotator naming mistake.

• skills/references/tracking.md — sv.ByteTrack setup, correct constructor param (track_activation_threshold, not confidence_threshold), correct update method (update_with_detections, not update), and filtering by tracker_id.

• skills/references/video.md — sv.process_video's show_progress param (the progress=True typo silently no-ops), VideoInfo attributes including fps as a float (not int, to avoid timestamp drift), and VideoSink for manual frame-by-frame writing.

• skills/references/utils.md — PolygonZone.trigger() for zone membership, LineZone counting with in_count/out_count, and sv.Color/sv.Color.from_hex/sv.ColorPalette.DEFAULT for styling.

Why

Agents frequently invent plausible-but-wrong supervision API surface (.filter(), BoundingBoxAnnotator, tracker.update(), progress=True) or reach for the wrong video-processing tool. This skill packages the correct patterns and the specific mistakes to avoid so agents get it right on the first try.

Testing

Docs-only change (markdown skill files, no source code touched). No tests affected.

[CLAassistant]

All committers have signed the CLA.

[Borda]

this was depercated

[Borda]

I think that shall be covered in llm.txt, and any agent shall be able to learn it from there