Adding Claude Code Skills

Skills are custom slash commands for Claude Code, defined as SKILL.md files. Each skill instructs Claude on how to perform a specific task using MCP tools.

Skill File Format

Skills live in .claude/skills/<name>/SKILL.md. Each file has YAML frontmatter and a Markdown body.

Frontmatter

---
name: my-skill
description: What this skill does (shown in Claude Code UI)
argument-hint: "<SYMBOL>"
user-invocable: true
allowed-tools: mcp__skopaq__get_quote mcp__skopaq__get_historical WebSearch
---

Field	Required	Description
`name`	Yes	Skill identifier (used as `/name` command)
`description`	Yes	Short description shown in Claude Code
`argument-hint`	No	Placeholder shown after the slash command
`user-invocable`	Yes	Set to `true` for user-facing skills
`allowed-tools`	No	Space-separated list of MCP tools the skill can use

Body

The Markdown body contains detailed instructions for Claude. Use $ARGUMENTS to reference what the user typed after the slash command.

# My Skill

You will perform X using data from MCP tools.

**Symbol to process: $ARGUMENTS**

## Step 1: Fetch Data
Call `mcp__skopaq__get_quote` with symbol=$ARGUMENTS.

## Step 2: Analyze
Based on the data, provide your analysis.

## Step 3: Present Results
Format as a table with these columns: ...

Example: Creating a /watchlist Skill

Step 1: Create the Directory

mkdir -p .claude/skills/watchlist

Step 2: Write the SKILL.md

---
name: watchlist
description: Monitor a list of stocks with live quotes and alerts.
argument-hint: "RELIANCE,TCS,INFY"
user-invocable: true
allowed-tools: mcp__skopaq__get_quote mcp__skopaq__get_historical
---

# Watchlist Monitor

Monitor the given stocks and highlight any with unusual activity.

**Stocks: $ARGUMENTS**

## Step 1: Fetch Quotes
For each symbol in the comma-separated list, call:
- `mcp__skopaq__get_quote` with symbol=SYMBOL

## Step 2: Analyze
For each stock, check:
- Change% > 2% (significant move)
- Volume unusually high
- Price near day high or day low

## Step 3: Present
Show a table:
| Symbol | LTP | Change% | Volume | Alert |

Highlight any stocks with unusual activity.

Step 3: Test

In Claude Code, type:

/watchlist RELIANCE,TCS,INFY,HDFCBANK

Existing Skills Reference

Skill	File	Tools Used
`/quote`	`.claude/skills/quote/SKILL.md`	`get_quote`
`/analyze`	`.claude/skills/analyze/SKILL.md`	`gather_all_analysis_data`, `get_quote`, `recall_agent_memories`, `check_safety`
`/scan`	`.claude/skills/scan/SKILL.md`	`get_quote`, `get_historical`, `WebSearch`
`/portfolio`	`.claude/skills/portfolio/SKILL.md`	`get_positions`, `get_holdings`, `get_funds`
`/trade`	`.claude/skills/trade/SKILL.md`	`get_quote`, `get_historical`, `get_funds`, `check_safety`, `place_order`
`/options`	`.claude/skills/options/SKILL.md`	`get_option_chain`, `suggest_option_trade`, `get_quote`

Best Practices

Use MCP Tools, Not Code

Skills should instruct Claude to use MCP tools, not write Python code:

# Good
Call `mcp__skopaq__get_quote` with symbol=$ARGUMENTS.

# Bad
Run this Python code:
```python
from skopaq.broker.client import INDstocksClient
...

Restrict Allowed Tools

Only list the tools the skill actually needs in allowed-tools. This prevents the skill from accidentally using destructive tools:

# Specific — only read-only tools
allowed-tools: mcp__skopaq__get_quote mcp__skopaq__get_historical

# Not recommended — allows everything
# (no allowed-tools field = all tools available)

Ask for Confirmation Before Actions

Any skill that modifies state (places orders, saves data) should ask for user confirmation:

## Step 4: Confirm
Present the trade plan and ask:
"Shall I execute this trade? (y/n)"

**NEVER execute without explicit user confirmation.**

Use $ARGUMENTS

The $ARGUMENTS variable contains everything the user typed after the slash command:

/analyze TCS         → $ARGUMENTS = "TCS"
/scan 10             → $ARGUMENTS = "10"
/options NIFTY PUT   → $ARGUMENTS = "NIFTY PUT"

Handle the case where $ARGUMENTS is empty by providing sensible defaults.

Structure with Clear Steps

Break the skill into numbered steps. Claude follows structured instructions better than freeform text:

## Step 1: Gather Data
...
## Step 2: Analyze
...
## Step 3: Present Results
...

Debugging Skills

If a skill is not working:

Check the file is at .claude/skills/<name>/SKILL.md (exact path matters)
Verify the YAML frontmatter is valid (no tabs, proper indentation)
Ensure user-invocable: true is set
Check that listed MCP tools exist and are spelled correctly
Restart Claude Code after adding or modifying skills