Context Diet

Optimizes context window usage for efficient Claude Code sessions.

When to Use

• "Too much context"
• "Slim down context"
• "Hitting token limits"
• "Session feels slow"
• "What should I keep in context?"

Context Budget

This Repo's Essential Files

Always Relevant (Keep Short Reference)

CLAUDE.md

Fetch When Needed

apps/web/app/api/_lib/

packages/db/src/schema.ts

packages/auth/src/index.ts

Skip (Already Known)

• Standard TypeScript patterns
• React/Next.js conventions
• Node.js fundamentals
• Common library usage

Procedure

Step 1: Assess Current Task

What's the user working on?

• Single file edit → Keep only that file
• Cross-package feature → Keep interface files, fetch impl
• Debugging → Keep error context, fetch related

Step 2: Identify Essential Context

For the current task, what's truly needed?

## Essential for This Task

1. {file1} - {why needed}
2. {file2} - {why needed}

## Can Fetch If Needed

- {file3} - Only if {condition}
- {file4} - Only if {condition}

## Skip

- {file5} - Already known / not relevant

Step 3: Context Shortlist

Generate a minimal file list:

## Context Shortlist

### Must Have (read fully)

- `{path}` - Active editing

### Reference Only (read signatures)

- `{path}` - Import types
- `{path}` - Check pattern

### Fetch If Blocked

- `{path}` - Implementation details

Step 4: Recommend Actions

## Recommendations

### To Reduce Context Now

1. {action} - saves ~{tokens}
2. {action} - saves ~{tokens}

### For Future Sessions

- Start with: "I'm working on {task}. Files: {shortlist}"
- Use subagents for exploration
- Split large tasks into focused sessions

Context Patterns

Bad (Context Bloat)

"Read the entire codebase"
"Show me all the packages"
"Give me everything about auth"

Good (Focused)

"I'm editing apps/web/app/api/users/route.ts"
"Show me the getCurrentUser function signature"
"What tables are in the schema?"

Token Estimates

Guardrails

• DO NOT recommend skipping CLAUDE.md
• Always include the file being actively edited
• Prefer interfaces over implementations
• Suggest subagents for broad exploration
• Be specific about what "fetch if needed" means