Facial recognition privacy protection — AI agent skill client. Detects faces locally (no data leaves the machine), then protects them from recognition systems using adversarial perturbations. Supports local CPU inference or GPU backend for higher quality.
Three tools with complete input schemas and descriptions. Naming follows verb_noun pattern (protect_face, detect_faces, analyze_face). Descriptions are adequate (100-150 chars) but lack actionable context for LLM selection. Parameter descriptions are present but sparse, 'mode' and 'tier' lack enum constraints or format guidance. Output schemas are not documented. Path validation is robust (traversal, symlink checks), but error messages are generic. No tool annotations (readOnlyHint, destructiveHint). Missing pagination, batch operations, and recovery guidance.
Analyze a face image's vulnerability to facial recognition systems. Requires a running goop-face-engine backend.
Detect faces in an image. Runs entirely locally — no data leaves your machine.
Protect a facial image from recognition systems. Face detection runs locally — the original image never leaves your machine. Only the detected face crop is processed (locally or sent to the GPU backend).
Parameter 'mode' and 'tier' lack enum constraints. Descriptions say 'stealth, balanced, maximum' and 'auto, local, server' but no formal enum in schema. LLMs may hallucinate invalid values like 'ultra' or 'gpu'.
Output schemas not documented. Tools return JSON strings but no schema is visible for what fields the response contains (e.g., does detect_faces return face_count, bounding_boxes, confidence scores?). LLMs cannot plan downstream calls.
No tool annotations. protect_face is destructive (writes files) but lacks destructiveHint. detect_faces and analyze_face are read-only but lack readOnlyHint. Agents cannot distinguish safe from unsafe operations.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | <=2025-11-25 | v2 |
Error handling is minimal. _validate_image_path raises ValueError/FileNotFoundError with brief messages. No recovery guidance (e.g., 'Image not found. Check the path or use detect_faces to list available images'). Agents cannot self-correct.
Parameter 'output_path' defaults to empty string, then resolves to config.output_dir or input stem. Behavior is implicit and not documented in the parameter description. LLMs cannot predict where files will be saved.