dance-lessons-coach/.vibe/skills/skill-creator/references/BEST_PRACTICES.md

# Skill Creation Best Practices

Based on the [Agent Skills Best Practices Guide](https://agentskills.io/skill-creation/best-practices)

## Core Principles

### Start from Real Expertise

Effective skills are grounded in real domain knowledge, not generic LLM knowledge. Feed project-specific context into the creation process.

**Sources of expertise:**
- Hands-on task completion with agent assistance
- Project artifacts (runbooks, API specs, schemas)
- Code review comments and issue trackers
- Version control history and patches
- Real-world failure cases and resolutions

### Refine with Real Execution

Test skills against real tasks and refine based on execution traces. Look for:
- False positives (skill activating when it shouldn't)
- Missed steps or edge cases
- Unproductive paths (agent trying multiple approaches)
- Context overload (too much irrelevant information)

## Context Management

### Spend Context Wisely

Every token in your skill competes for attention in the agent's context window.

**Add what the agent lacks:**
- Project-specific conventions
- Domain-specific procedures
- Non-obvious edge cases
- Specific tools/APIs to use

**Omit what the agent knows:**
- Basic concepts (what is a PDF, HTTP, database)
- Generic best practices
- Obvious implementation details

### Design Coherent Units

Skills should encapsulate a coherent unit of work that composes well with others:
- **Too narrow**: Multiple skills needed for one task
- **Too broad**: Hard to activate precisely
- **Just right**: One skill handles one class of problems

### Aim for Moderate Detail

Concise, stepwise guidance with working examples outperforms exhaustive documentation.

## Instruction Patterns

### Gotchas Sections

List environment-specific facts that defy reasonable assumptions:

```markdown
## Gotchas

- The `users` table uses soft deletes (add `WHERE deleted_at IS NULL`)
- User ID is `user_id` in DB, `uid` in auth service, `accountId` in billing API
- `/health` returns 200 even when DB is down (use `/ready` for full health check)
```

### Templates for Output Format

Provide concrete format examples rather than prose descriptions:

```markdown
## Report Structure

```markdown
# [Analysis Title]

## Executive Summary
[One-paragraph overview]

## Key Findings
- Finding 1 with data
- Finding 2 with data

## Recommendations
1. Actionable recommendation
2. Actionable recommendation
```
```

### Checklists for Multi-Step Workflows

```markdown
## Deployment Workflow

Progress:
- [ ] Step 1: Run tests
- [ ] Step 2: Build artifacts
- [ ] Step 3: Validate configuration
- [ ] Step 4: Deploy to staging
- [ ] Step 5: Run smoke tests
```

### Validation Loops

```markdown
## Code Review Process

1. Make changes
2. Run linter: `npm run lint`
3. If errors: fix and re-lint
4. Run tests: `npm test`
5. If failures: fix and re-test
6. Only commit when all checks pass
```

### Plan-Validate-Execute Pattern

```markdown
## Database Migration

1. Generate migration plan: `migrate plan`
2. Review plan against schema
3. Validate: `migrate validate`
4. If invalid: revise and re-validate
5. Execute: `migrate apply`
```

## Control Calibration

### Match Specificity to Fragility

**Give freedom** when multiple approaches are valid:
```markdown
## Code Review

Check for:
- SQL injection vulnerabilities
- Proper authentication
- Race conditions
- PII in error messages
```

**Be prescriptive** when operations are fragile:
```markdown
## Database Migration

Run exactly:
```bash
python scripts/migrate.py --verify --backup
```

Do not modify this command.
```

### Provide Defaults, Not Menus

```markdown
<!-- Avoid -->
You can use pypdf, pdfplumber, PyMuPDF, or pdf2image...

<!-- Better -->
Use pdfplumber for text extraction:

```python
import pdfplumber
```

For scanned PDFs, use pdf2image with pytesseract.
```

### Favor Procedures Over Declarations

Teach *how to approach* problems, not *what to produce*:

```markdown
<!-- Avoid -->
Join orders to customers on customer_id, filter region='EMEA', sum amount.

<!-- Better -->
1. Read schema from references/schema.yaml
2. Join tables using _id foreign keys
3. Apply user filters as WHERE clauses
4. Aggregate and format as markdown table
```

## Progressive Disclosure

### Keep SKILL.md Under 500 Lines

Move detailed reference material to separate files in `references/`:

```
skill-name/
├── SKILL.md              # Core instructions (<500 lines)
├── references/
│   ├── api-spec.md       # Detailed API documentation
│   ├── error-codes.md    # Error code reference
│   └── schemas/          # Data schemas
└── scripts/
    └── validate.sh       # Validation script
```

### Load Context on Demand

Tell the agent *when* to load reference files:

```markdown
Read references/api-errors.md if the API returns non-200 status.
```

## Skill Structure Checklist

### Required Elements
- [ ] `SKILL.md` with valid YAML frontmatter
- [ ] `name` field (lowercase alphanumeric + hyphens, 1-64 chars)
- [ ] `description` field (1-1024 chars, specific about what/when)
- [ ] Clear instructions in Markdown body

### Recommended Elements
- [ ] `license` field
- [ ] `metadata` with author/version
- [ ] `scripts/` directory for reusable code
- [ ] `references/` directory for detailed docs
- [ ] `assets/` directory for templates/resources

### Validation Checklist
- [ ] Skill name matches directory name (underscores → hyphens)
- [ ] Description is specific and actionable
- [ ] Instructions focus on what agent wouldn't know
- [ ] Gotchas section for non-obvious issues
- [ ] Examples provided for key workflows
- [ ] Progressive disclosure used for large skills
- [ ] Validation loops for critical operations

## Iteration Process

1. **Create initial draft** from real expertise
2. **Test against real tasks** 
3. **Review execution traces** for inefficiencies
4. **Refine instructions** based on observations
5. **Add gotchas** from corrections made
6. **Validate** with skill_creator
7. **Repeat** until performance is satisfactory

## Common Anti-Patterns

### ❌ Generic Advice
```markdown
Handle errors appropriately and follow best practices.
```

### ✅ Specific Guidance
```markdown
Check for HTTP 429 errors and implement exponential backoff:

```python
import time
import requests

for attempt in range(5):
    try:
        response = requests.get(url, timeout=10)
        break
    except requests.HTTPError as e:
        if e.response.status_code == 429:
            time.sleep(2 ** attempt)
        else:
            raise
```
```

### ❌ Overly Broad Scope
```markdown
This skill handles all database operations including queries, migrations, backups, and administration.
```

### ✅ Focused Scope
```markdown
This skill handles database query optimization for read-heavy workloads on PostgreSQL 14+.
```

### ❌ Menu of Options
```markdown
You can use any of these libraries: pandas, numpy, polars, or vaex.
```

### ✅ Clear Default
```markdown
Use polars for DataFrame operations:

```python
import polars as pl
df = pl.read_csv("data.csv")
```

For pandas compatibility, use the .to_pandas() method.
```