Creating Skills¶

Learn how to build custom Claude Code skills for the RQF-ML system.

Quick Start¶

Use the /skill-create meta-skill:

/skill-create my-skill "Description of what it does"

Or create manually:

mkdir -p .claude/skills/my-skill
# Create SKILL.md with frontmatter and instructions

Skill Structure¶

Minimal (Single File)¶

.claude/skills/my-skill/
└── SKILL.md

Standard (With Templates)¶

.claude/skills/my-skill/
├── SKILL.md
└── templates/
    └── output.md

Complex (Full Suite)¶

.claude/skills/my-skill/
├── SKILL.md
├── templates/
│   ├── report.md
│   └── checklist.md
├── examples/
│   └── example_usage.md
└── scripts/
    └── helper.py

SKILL.md Format¶

---
name: my-skill
description: "USE FOR: X when Y. NOT FOR: Z"
disable-model-invocation: true  # Manual only
argument-hint: [file] [--flag]
allowed-tools: Read, Grep, Glob, Write
---

# My Skill

One sentence describing what this skill does.

## Usage

## Arguments
| Arg | Required | Description |
|-----|----------|-------------|
| `$0` | Yes | First argument |
| `--flag` | No | Optional flag |

## Workflow

### Step 1: Do Something
Instructions for step 1.

**Check:** How to verify this step.

### Step 2: Do More
Instructions for step 2.

## Output
- **Location:** Where results go
- **Format:** What format

## Error Handling
| Error | Cause | Solution |
|-------|-------|----------|
| Error 1 | Why | How to fix |

Frontmatter Reference¶

Required¶

Recommended¶

Optional¶

Skill Types¶

Action Skill (Side Effects)¶

---
name: deploy
description: "USE FOR: Deploying to production"
disable-model-invocation: true  # IMPORTANT!
allowed-tools: Read, Grep, Bash(*)
---

Analysis Skill (Read-Only)¶

---
name: analyze-code
description: "USE FOR: Code analysis when asked"
allowed-tools: Read, Grep, Glob
---

Generator Skill (Creates Files)¶

---
name: generate-report
description: "USE FOR: Creating reports"
allowed-tools: Read, Grep, Glob, Write
---

Research Skill (Isolated)¶

---
name: deep-research
description: "USE FOR: Deep codebase exploration"
context: fork
agent: Explore
---

Reference Skill (Background)¶

---
name: coding-standards
description: "Reference for coding standards"
user-invocable: false  # Claude uses internally
---

Arguments¶

Basic Substitution¶

---
argument-hint: [filename]
---

Analyze the file at $ARGUMENTS

Individual Arguments¶

---
argument-hint: [source] [target]
---

Copy from $0 to $1

Dynamic Context¶

---
name: pr-review
---

## PR Context
- Diff: !`gh pr diff`
- Comments: !`gh pr view --comments`

Review the above...

Quality Checklist¶

Before deploying a skill:

• Description starts with "USE FOR:"
• Includes "NOT FOR:" anti-patterns
• Every step has verification
• Error handling for top 3 failures
• At least 2 usage examples
• Tool restrictions are minimal
• No assumed knowledge

Quality Scoring¶

Grades:

• 90-100: Production ready
• 70-89: Good, minor improvements
• 50-69: Needs work
• <50: Major revision needed

Examples¶

Example 1: Data Refresh Skill¶

---
name: data-refresh
description: "USE FOR: Refreshing market data from FirstRate"
disable-model-invocation: true
argument-hint: [symbol]
allowed-tools: Read, Write, Bash(python *)
---

# Data Refresh

Fetch latest market data and validate quality.

## Usage

## Workflow

### Step 1: Fetch Data
```bash
python scripts/fetch_data.py --symbol $0

Step 2: Validate¶

python scripts/validate_data.py data/raw/$0_*.csv

Step 3: Archive¶

Move validated data to data/processed/

Output¶

• data/raw/$0_[date].csv
• data/processed/$0_validated.csv

  ### Example 2: Code Review Skill
  
  ```yaml
  ---
  name: review-pine
  description: "USE FOR: Reviewing Pine changes before commit"
  allowed-tools: Read, Grep, Glob
  ---
  
  # Pine Code Review
  
  Review PineScript changes for quality and compliance.
  
  ## Checklist
  
  ### Non-Repaint
  - [ ] All signals use `barstate.isconfirmed`
  - [ ] No `request.security` with lookahead
  - [ ] No `barstate.isrealtime` for signals
  
  ### State Machine
  - [ ] All transitions defined
  - [ ] Timeout handling present
  - [ ] Reset conditions clear
  
  ### Memory
  - [ ] Array cleanup exists
  - [ ] Drawing limits respected
  

Testing Skills¶

1. Manual Test: /my-skill test-args
2. Edge Cases: Test with invalid input
3. Auto-Invocation: Verify Claude triggers appropriately

Debugging¶

# Check skill is loaded
/skills-list

# Verify file exists
ls -la .claude/skills/my-skill/SKILL.md

# Check frontmatter syntax
cat .claude/skills/my-skill/SKILL.md | head -20