ml-cv-specialist

I've been digging through computer vision skills lately, and your approach to grounding ML/CV concepts in practical Claude workflows is refreshing—most skills either go too theoretical or skip the implementation entirely, but you managed to thread that needle pretty well.

Links:

• GitHub
• Your agentic skill in the SkillzWave marketplace

TL;DR

You're at 90/100, solid A-grade territory. This is based on Anthropic's best practices for skill design. Your strongest area is Writing Style (9/10)—the content is dense and appropriately technical without fluff. The weakest link is Spec Compliance (12/15), which is fixable with a couple of specific additions.

What's Working Well

• Layered architecture is chef's kiss. SKILL.md gives the overview, model-catalog.md handles the deep benchmarks—clean separation that respects reader attention. The progressive disclosure structure is exactly what Claude needs.
• Your decision trees actually work. The "API vs. Self-Hosted" framework and "I need to classify images" tables make this actionable. Not just theoretical—someone can actually use this to pick a model.
• Trigger phrases hit the mark. "Designing machine learning systems" and "computer vision pipelines" are exactly what people search for. You nailed the discoverability language.
• Objectivity throughout. Zero marketing fluff, all specifications and trade-offs. The cost/latency/accuracy tables are the kind of thing that actually moves decisions forward.

The Big One

Missing reference file is breaking your PDA structure. Line 382 references inference-patterns.md that doesn't exist. This isn't just a dead link—it signals incomplete architecture to anyone reading the layered structure.

Fix: Either create the file with the promised architecture patterns (batch inference, streaming, real-time), or remove the reference and fold the essential patterns into SKILL.md. If you go with the second option, add 2-3 concrete implementation examples (YOLOv8 setup, model serving flow) to fill that gap. This'll bump you +2 points easily.

Other Things Worth Fixing

1. Add a table of contents. SKILL.md is 383 lines without a TOC. At that length, readers need quick navigation. Add a ## Contents section right after your metadata with links to main sections. Similar issue in model-catalog.md (259 lines). Low effort, +2 points.

2. Switch "Your Situation" tables to neutral phrasing. Lines 232-259 use second-person perspective; reframe as "Image Classification Selection" with neutral "Requirement" columns instead. Keeps the utility, improves objectivity. +1 point.

3. Flesh out feedback loops. Your Monitoring section (lines 325-350) lists metrics but skips the run→check→fix workflow. Add one concrete example: "1. Deploy model → 2. Monitor P95 latency → 3. If >2x baseline: check GPU utilization → 4. Scale or optimize batch size → 5. Re-measure." Shows how to actually act on the data. +1 point.

Quick Wins

• Create or remove that missing inference-patterns.md reference (+2)
• Add TOC to SKILL.md and model-catalog.md (+2)
• Switch "Your Situation" tables to neutral voice (+1)
• Add one complete feedback loop example to Monitoring (+1)

Four concrete changes, realistic time investment, and you're at 96/100 territory.

Checkout your skill here: SkillzWave.ai | SpillWave We have an agentic skill installer that installs skills in 14+ coding agent platforms. Check out this guide on how to improve your agentic skills.