Build a Context Engineering Strategy

Context engineering is about making sure Claude has exactly the information it needs, and nothing it doesn’t.

The Three Layers

Layer 1: Always loaded (CLAUDE.md) Keep this small. Only rules that apply to every single task.

# CLAUDE.md
Use TypeScript. Run `bun test` to verify. Follow existing patterns.

Layer 2: Conditionally loaded (.claude/rules/) Rules that apply only in certain directories or file types.

# .claude/rules/api.md
All API routes go in src/routes/. Use Zod for request validation.
Return consistent error shapes: { error: string, code: number }.

Layer 3: On-demand (skills, @ references) Heavy context loaded only when needed.

@src/db/schema.ts implement a new users table following this schema pattern

Minimize Token Waste

• Use .claudeignore to exclude node_modules, dist, build artifacts, and large generated files
• Keep CLAUDE.md under 200 lines (aim for under 50)
• Use .claude/rules/ for directory-specific instructions instead of bloating CLAUDE.md
• Reference files with @ instead of pasting their contents

Session Habits

• Start with /clear between unrelated tasks
• Use /compact when context gets long
• Front-load important context in your first message
• Tell Claude which files matter: “Focus on src/auth/ for this task”

Measure Your Context

Run /cost periodically. If you’re burning through context fast, check whether Claude is reading files it doesn’t need. Add those patterns to .claudeignore.