Custom Rules in Claude Code CLAUDE.md | Generated by AI

Home 2026.03

Question: How do you make Claude Code follow custom rules — should you add them to CLAUDE.md or use skills?

Answer:

Great question. Claude Code has several mechanisms for injecting persistent instructions. Here’s a comprehensive breakdown:

1. CLAUDE.md — The Primary Method

This is the most recommended and native way to give Claude Code persistent instructions.

How it works:

File locations (all are read, in order):

Location Scope
~/.claude/CLAUDE.md Global — applies to ALL projects
<project-root>/CLAUDE.md Project-level — applies to this repo
<subdir>/CLAUDE.md Subdirectory-level — applies when working in that folder

Example ~/.claude/CLAUDE.md for your formatting rules:

## Response Format Rules

Always respond in this exact format:

**Question:** [rephrased and grammar-fixed user question]

**Answer:**
[full answer in clean Markdown, always in English]

**References:**
[only show if there is at least one, use bullet list with [url](url) format]

- No Grok render tags
- Do not show images, prefer text only
- Give comprehensive explanations for exam questions
- Always answer in English, even if the question is in Chinese
- Always search the web before answering

2. Project-level CLAUDE.md — For Team/Repo Rules

Place a CLAUDE.md in your project root for rules specific to that codebase:

## Project Conventions
- Use TypeScript strict mode
- All functions must have JSDoc comments
- Run `npm test` before committing

This is version-controlled and shared with your team automatically.

3. Skills (/mnt/skills/) — For Claude.ai Web, Not Claude Code

Skills (like the /mnt/skills/ system you see in Claude.ai) are a feature of the Claude.ai web interface, not Claude Code CLI. They are pre-loaded instruction files injected by Anthropic’s platform.

Claude Code does NOT support the skills system. Don’t confuse the two.

4. System Prompt via API / --system-prompt flag

If you’re calling Claude Code programmatically or via the SDK, you can pass a system prompt directly:

claude --system-prompt "Always answer in English. Follow this format: ..."

Or in the SDK:

const claude = new Anthropic();
// pass system prompt in your API call

5. /memory Command (Runtime)

Inside a Claude Code session, you can use:

/memory

This lets you add or edit memory that persists across sessions — similar to CLAUDE.md but managed interactively.

Best Practice Summary

Goal Method
Rules for ALL projects globally ~/.claude/CLAUDE.md
Rules for one specific project <project>/CLAUDE.md
Rules for a subfolder <subdir>/CLAUDE.md
One-off session rules Type instructions inline or use /memory
Programmatic/API usage --system-prompt flag or SDK system param

Recommendation for Your Use Case

Since your formatting rules (English-only, specific output structure, always search) are personal preferences that should apply everywhere, put them in:

~/.claude/CLAUDE.md

This is the global CLAUDE.md and will apply to every project you work on with Claude Code automatically — no per-project setup needed.

References:

Back Donate