Expert Skill Creation
Transform expert knowledge into production-ready skills that combine domain expertise with system-specific understanding.
Why Skills Fail in Production
AI assistants fail not because they lack intelligence, but because they lack:
- Domain Expertise — Industry-specific rules, edge cases, unwritten conventions
- Ontology Understanding — How YOUR systems, data structures, and workflows actually work
Both are required. Domain knowledge without system context produces generic output. System knowledge without domain expertise produces structurally correct but semantically wrong results.
Workflow
Assess → Discover (Expertise + Ontology) → Design → Create → Refine → ShipQuick Assessment
Create a skill when:
- Used 3+ times (or will be)
- Follows consistent procedure
- Saves >300 tokens per use
- Requires specialized knowledge not in Claude's training
- Must produce trusted output (not "close enough")
Don't create for: one-time tasks, basic knowledge Claude already has, rapidly changing content.
Discovery: Two Streams
Stream 1: Domain Expertise
Deep knowledge that transcends any specific company:
- Industry standards and their versions
- Professional conventions and best practices
- Edge cases only practitioners know
- Validation rules from specifications
*Example (LEDES validation):* LEDES 98B vs XML 2.0 formats, UTBMS code taxonomy, date format requirements, required vs optional fields.
Stream 2: Ontology Understanding
How the skill maps to specific systems and organizations:
- Company-specific policies and constraints
- Data structures and identifiers unique to the system
- Cross-references between entities (timekeepers → IDs → rates)
- Workflow states and transitions
*Example (LEDES validation):* Firm-specific timekeeper codes, matter numbering conventions, approved billing rates, outside counsel guideline requirements.
Discovery Questions
When starting, I'll ask about:
- Domain & Purpose — What problem? What industry standards apply?
- Ontology Requirements — What system-specific structures must the skill understand?
- Content Source — Conversation, docs, specifications, or files to distill from?
- Automation Potential — What can be deterministic (scripts)? What needs interpretation (LLM)?
- Complexity Level — Simple (SKILL.md only), Enhanced (+scripts), or Full (+resources)?
Skill Architecture
skill-name/
├── SKILL.md # Layer 1: Core (300-500 tokens)
├── scripts/ # Layer 0: Automation (0 tokens to run)
│ └── validate.py
└── resources/ # Layer 2: Details (loaded selectively)
└── ADVANCED.mdLayer 0 (Scripts): Free execution, structured JSON output Layer 1 (SKILL.md): Loaded when triggered - keep lean Layer 2 (Resources): Fetched only when specific section needed
Token Optimization
| Technique | Instead of | Do this | Savings |
|---|---|---|---|
| Scripts | 500 tokens explaining validation | python scripts/validate.py | ~450 tokens |
| Reference | Inline schema (200 tokens) | Link to resources/schema.json | ~185 tokens |
| Layer 2 | Everything in SKILL.md | Link to resources/ADVANCED.md | ~750 tokens |
Description Formula
<Action> <Object> for <Purpose>. Use when <Trigger>.
Example: "Validate billing data for system migration. Use before importing invoices."
Shipping
When content is finalized:
python scripts/package_skill.py skill-name 1.0Creates skill-name-v1.0.zip with:
- DIRECTORY_STRUCTURE.txt (auto-generated)
- README.md with deployment instructions
- All skill files properly organized
Templates & Examples
See resources/templates/ for:
- Minimal skill template
- Enhanced skill template
- Script template
See resources/examples/ for domain-specific patterns.
Quality Checklist
Before shipping:
- Description <30 tokens
- SKILL.md <500 tokens (Layer 1)
- Scripts for deterministic operations
- Advanced content in resources/ (Layer 2)
- Version in frontmatter
- All referenced files exist
Version: 2.2 | Target: <500 tokens Layer 1