Conciseness

Every section is imperative and load-bearing — a backend-selection table, minimal quick starts with full commands, and gotchas that carry non-obvious information ('some ops fall back to CPU; validate latency on a real device'). No concept Claude already knows is explained; nothing reads as padding. Matches the 5 anchor 'every token earns its place'.

Actionability

Commands are copy-paste executable with concrete flags: 'python -m openmed.mlx.convert --model … --output ./mlx_ner --quantize 8', 'pip install "openmed[mlx]"', 'convert(…, compute_units="cpuAndNeuralEngine", compute_precision="float16", quantize="int8")', and 'include_webgpu=True, opset=18'. The only placeholders ('OpenMed/<some-ner-model>', prompt="...") are user-supplied parameters, not missing detail, so this stays at 5 rather than 4's 'minor gaps'.

Workflow Clarity

The Workflow section gives a clear pick → convert → quantize → run → verify sequence with explicit validation checkpoints ('certify recall' via '--eval-suite writes a recall-delta report', 'gate on leakage, not just F1', verify against full precision before shipping). It falls short of the 5 anchor only because there is no explicit error-recovery loop — what to do when the quantized model fails the recall gate is left implicit — while clearly exceeding 3, which requires missing or implicit checkpoints.

Progressive Disclosure

A single well-organized SKILL.md with clean section headers, a comparison table, and a 'Standards & references' section whose links (MLX/Core ML Tools/ONNX/transformers.js docs, 'openmed/mlx/convert.py' source paths) are one level deep and clearly signaled. It is not the 5 anchor because at ~165 lines the three per-backend quick-starts and gotcha lists are inlined where per-backend reference files would keep the overview leaner; it is well above 3, which requires buried references or structureless inline bulk.

Specificity

Quotes multiple specific concrete actions: 'Run OpenMed models fully on-device with the MLX …, CoreML …, or ONNX/WebGPU … backends, including convert-quantize-run workflows' plus 'the convert() functions and python -m convert CLIs, quantization, loading a local MLX artifact through analyze_text, OpenMedMLXLanguageModel/generate_text, and the on-device-only PHI guarantee'. Coverage is comprehensive across all three backends with no gaps; this clearly matches the 5 anchor rather than 4, which would require minor coverage gaps.

Completeness

It explicitly answers both: what — 'Run OpenMed models fully on-device with the MLX/CoreML/ONNX backends, including convert-quantize-run workflows' — and when — the explicit 'Use when…' clause with concrete trigger phrases. This is the 5 anchor verbatim in structure; 4 would require the 'when' to be less explicit.

Trigger Term Quality

'Use when the user wants to deploy OpenMed at the edge, run NER/de-id on Apple Silicon, target iPhone/iPad/Mac, export to ONNX or WebGPU, quantize a clinical model to int8/4-bit, run with no network, or pick between MLX/CoreML/ONNX' covers natural phrasings and synonyms comprehensively (edge, on-device, no network, device names, format names, quantization levels). It matches the 5 anchor's 'including synonyms' breadth; nothing common is missing.

Distinctiveness Conflict Risk

The 'fully on-device / at the edge' framing plus backend-specific triggers (Apple Silicon, iPhone/iPad, WebGPU, no network) carve a clear niche distinct from hosted serving, MCP tooling, or model-selection skills. Conflict risk is minimal; it does not fall to 4, which presumes overlap with closely related skills.