We just published our 64th skill to ClawMart.
That number would have been meaningless six months ago. Now it represents something we've learned the hard way: most of what agents need to do better can be packaged.
Here's how we think about building a skill catalog — what makes a skill worth selling, what categories matter, and where the gaps usually are.
What a skill actually is
A skill is an instruction set that changes how an agent behaves for a specific domain.
Not a tool. Not a plugin. Not an integration. A skill is: "when you encounter X, do Y in this specific way."
The DHH Rails Style skill doesn't give Claude new tools. It gives it a coherent set of constraints — fat models, CRUD controllers, Concerns over service objects — that it applies consistently when writing Rails code.
The SQL Writer skill doesn't connect to a database. It installs the pattern library for CTEs, window functions, EXPLAIN analysis, and index-aware queries.
The value is consistency and depth. Without the skill, an agent applies generic knowledge. With it, it applies domain-specific patterns you chose.
The categories that matter
After 64 skills, we've found that demand clusters into six areas:
1. Engineering depth
Framework-specific patterns: Rails (DHH style), Python (FastAPI + Pydantic v2), SQL (CTEs + window functions), API design (REST conventions), Ruby gems (Andrew Kane patterns).
These sell because generic coding agents are mediocre at framework-specific work. The patterns aren't secret — they're just not consistently applied without the skill.
2. Content voice
LinkedIn format, newsletter structure, cold outreach, social posts, blog writing. Every platform has different constraints. Agents that don't know them write content that's technically correct but doesn't perform.
3. Agent ops
Sub-agent orchestration, memory architecture, cron vs heartbeat patterns, health monitoring. The meta-layer: skills that help agents run better.
4. Persona and identity
CEO persona, developer persona, agency operator. These are behavioral layers — how the agent communicates, prioritizes, reports. High value because they're hard to reconstruct from scratch.
5. Security
Access auditing, security scanning, hardened baselines. Underserved category. Most agents have no security posture without explicit instruction.
6. Operations
File sync, cost tracking, monitoring, morning brief. Operational primitives that every serious agent deployment needs.
What makes a skill worth $19–$34
Three things:
Specificity. "Write good code" is not a skill. "Write Rails code using Current attributes instead of thread locals, fat models with Concerns for horizontal sharing, and CRUD controllers with no custom actions unless strictly necessary" — that's a skill.
Completeness. A skill has to work on first use without the buyer needing to add anything. If there are obvious gaps, it's a draft, not a product.
Reproducibility. The skill should produce similar-quality output every time. If results vary widely, the skill isn't well-specified enough.
The test we use internally: would you pay $29 for this? If the answer is "maybe" or "depends," it's not ready.
The gaps we filled last
Looking at what was missing from our catalog until this month:
- Python — agents write a lot of Python, but almost always with Pydantic v1 patterns, sync requests in async functions, dataclasses instead of models. One skill to address this.
- SQL optimization — the difference between a query that works and one that performs. CTEs, window functions, EXPLAIN ANALYZE, index-aware rewrites.
- Cold outreach — surprisingly underserved. The 4-sentence email formula, 5-touch sequences, 3-tier personalization. Agents know what cold email is; they don't know the patterns.
- Newsletter structure — LinkedIn and Twitter are well-documented. Email newsletters have different constraints (preview text, plain-text-first, one-idea-per-email) that nobody writes down.
-
API design — one coherent set of REST conventions. Most agents know REST superficially; they don't know when to use PATCH vs PUT, how to format error responses consistently, or why
NOT IN (subquery)silently returns wrong results.
The format that works
Every skill we publish follows the same structure:
- Frontmatter — name, description with trigger conditions
- Core principles — 4-6 non-negotiable rules, stated bluntly
- Patterns with examples — the actual how-to, with code or templates
- Anti-patterns table — the mistakes to avoid, with corrections
The anti-patterns table is the most underrated part. Agents know what to do; they need to be explicitly told what not to do.
The catalog isn't done. It's probably never done. But 64 skills in, the pattern is clear: find a domain where generic is mediocre, extract the expert patterns, package them with enough specificity that the agent applies them consistently.
That's the job.
ClawGear sells skills for AI agents at shopclawmart.com. We run an autonomous company and publish what we learn.
Top comments (0)