Claude Code Financial Cheatsheet

Overview

A comprehensive enterprise reference for using Claude Code in quantitative finance workflows. Master CLI execution flags, slash commands, MCP architecture, sub-agent patterns, and security guardrails for building institutional-grade algorithmic trading systems.

1. CLI Execution Flags

Claude Code can be initialized with various flags to configure its environment.

• --model: Specify the underlying Claude model (e.g. claude-3-7-sonnet-20250219).
• --tools: Define which MCP (Model Context Protocol) servers to expose to the agent.
• --read-only: Force the agent into a read-only mode to prevent accidental deletion of backtest logs.

2. Slash Commands

• /goal: Sets a long-running goal for the agent, preventing it from yielding until the objective is complete. Useful for running extensive backtests.
• /schedule: Schedules a background cron job for the agent to wake up and check market conditions.
• /learn: Instructs the agent to persist a newly learned market structure quirk to its permanent memory.

3. Sub-Agent Patterns

In complex quantitative workflows, you should avoid monolithic agents. Instead, use a swarm of sub-agents:

1. Data Agent: Specialized in fetching and cleaning time-series data via SQL.
2. Quant Agent: Specialized in fitting mathematical models (e.g., GARCH) to the data.
3. Execution Agent: Specialized in taking the model's signals and generating FIX protocol messages.

4. Security Guardrails

When giving an LLM access to a terminal, security is paramount.

• Always run Claude Code inside an isolated Docker container with zero network access to production trading servers.
• Use the accidental-data-loss-prevention skill to require human confirmation before executing DROP TABLE or rm -rf commands.