✍️ Flawless description — frontend-design
Signal #1. A sample description says what the skill does, when to use it and list concrete triggers. frontend-design (488.299 installs, anthropics/skills) is the textbook case: it lists task examples and the desired effect, without making the agent guess.
real description (frontend-design · anthropics/skills · 488k):
Create distinctive, production-grade frontend interfaces with high design quality. Use this skill when the user asks to build web components, pages, artifacts, posters, or applications (examples include websites, landing pages, dashboards, React components, HTML/CSS layouts, or when styling/beautifying any web UI). Generates creative, polished code and UI design that avoids generic AI aesthetics.
✓ What makes it exemplary
- ✓Starts with the result: “production-grade frontend interfaces”.
- ✓"Use this skill when..." — the when explicit.
- ✓Parentheses with examples: websites, dashboards, React, HTML/CSS.
- ✓Stated differentiator: "avoids generic AI aesthetics".
✗ What a weak one would do
- ✗"Helps with frontend." — no when, no examples.
- ✗No trigger words the agent recognizes in the task.
- ✗No differentiator: why this one and not another?
"Pushy" variant: supabase
supabase (99k) opens by practically forcing it to trigger: "Use when doing ANY task involving Supabase. Triggers:..." — explicitly states the triggers and prompts the agent to activate it. This counters under-triggering (Claude tends not to activate it enough).
Lesson: the description isn’t a polite summary—it’s a routing instruction. Be specific in the when and list triggers.
🎯 Perfect atomic scope — vercel-react-best-practices
One responsibility, done exceptionally well. vercel-react-best-practices (443.261, vercel-labs/agent-skills) doesn’t try to cover "everything about web": it covers only React best practices. Because it’s atomic, it triggers precisely and works alongside web-design-guidelines and frontend-design without overlap.
| Atomic skill | Does ONE thing | Installs |
|---|---|---|
| vercel-react-best-practices | React patterns | 443k |
| test-driven-development | test before code | 107k |
| web-design-guidelines | layout & accessibility | 358k |
test-driven-development: process scope
test-driven-development (107k, obra/superpowers) is the purest example of atomicity: a single rule — "write the test before the code." It doesn’t mention a framework or a language. One responsibility, one trigger ("when implementing a feature or bug fix").
Because it’s so narrow, it never triggers incorrectly and always composes with whichever stack skill is active.
💡 Atomic scope test
Can you describe what the skill does in one sentence without using “and”/“also”? If so, it’s atomic. If the sentence has three clauses, they’re probably three skills.
🏛️ Strong, well-maintained source — microsoft/azure suite
Who publishes and maintains it matters as much as the content. The suite microsoft/azure-skills — foundry, ai, deploy, diagnostics, prepare, all ~358–360k installs — it’s the picture of a strong source: an official vendor, a cohesive suite, and ongoing maintenance that keeps pace with the platform itself.
Official vendor
Microsoft writes about Azure. Whoever operates the platform knows the edge cases no one documents.
Cohesive suite, not a monolith
Five atomic skills that work together — a sign that the source understands skill design, not just the product.
Ongoing maintenance
Because they install via symlink, npx skills update pulls improvements from the source. An active vendor means the skill is always up to date.
Other model sources
vercel-labs powers find-skills (1,8M), vercel-react-best-practices (443k), and web-design-guidelines (358k) — an entire portfolio of reference skills.
supabase e neondatabase write the banks’ own practices; stripe/ai, the payment ones. Vendor-backed is the highest confidence level.
💡 Source heuristic
Ask: “Is this skill about a product, and is its author the owner of that product?” If so, it’s vendor-backed—the strongest source signal there is.
🔍 Clear Discovery — find-skills
The most installed skill in the entire catalog: find-skills, 1.802.925 installs (vercel-labs/skills). It wins on one specific signal — discovery: its purpose (finding and suggesting other skills) is instantly obvious, which is why it becomes the first one everyone installs.
✓ Why it becomes a discovery model
- ✓Self-explanatory name: find-skills says everything.
- ✓Single purpose: help find the right skill.
- ✓Entry point: installed before anything else.
- ✓A strong source (vercel-labs) reinforces the adoption.
The scale of the signal
find-skills alone (1.8M) surpasses entire groups. The top 100 skills = 43.7% of all installs; only 0.3% (131 skills) exceed 100k. A clear name + purpose is what propels a skill into this elite group.
Clear discovery isn’t just marketing: it’s about the agent being able to find and choose the skill at the right moment.
clear discovery = a name + description that explain themselves:
find-skills → "encontra a skill certa" (óbvio) skill-creator → "cria skills" (óbvio) xyz-helper-v2-final → "???" (ninguém instala)
💡 Tip
If the name needs explaining, discovery has already failed. find-skills, skill-creator, frontend-design — they all say what they do without opening SKILL.md.
🧩 When a skill gets several things right — skill-creator
The models above excel at a signal. The real references get several. skill-creator (246k, anthropics) has a clear description + focused scope (creating/evaluating skills) + official Anthropic source + obvious discoverability from its name. It’s the meta-skill you’ll study in Track 3.
| Skill | Descr. | Scope | Source | Discov. |
|---|---|---|---|---|
| skill-creator | ✓ | ✓ | ✓ | ✓ |
| frontend-design | ✓ | ✓ | ✓ | ✓ |
| find-skills | ✓ | ✓ | ✓ | ✓ |
| generic "sql-helper" | ✗ | ✗ | ✗ | ✗ |
The compounding effect
Getting all four signals right at the same time doesn’t add—it multiplies. That’s why these three are among the most installed in the catalog: the agent finds them, understands them, triggers them correctly, and trusts the source. The four signals reinforce each other.
📋 Top performers by signal
The module’s reference table. For each quality signal, the skill to emulate — with real installs and sources. Use this table as a benchmark when judging (or creating) a skill.
| Signal | Skill template | Source | Installs |
|---|---|---|---|
| Description | frontend-design | anthropics | 488k |
| "Pushy" description | supabase | supabase | 99k |
| Atomic scope | vercel-react-best-practices | vercel-labs | 443k |
| Process scope | test-driven-development | obra/superpowers | 107k |
| Strong/maintained source | azure-skills (suite) | microsoft | ~358k |
| Clear discovery | find-skills | vercel-labs | 1,8M |
| Gets everything right | skill-creator | anthropics | 246k |
💡 How to use the table
When evaluating a new skill, compare it with the champion for the signal that matters most to you. When creating your own (module 2.4), copy the champion's pattern — don't invent one from scratch.
✅ Module Summary
Next:
Module 2.4 — 🛠️ How to Create a Skill That Looks (and Is) High Quality. Now that you’ve seen the examples, let’s imitate them in practice: a description that triggers, atomic scope, and a polishing checklist.