"Token Budget Management: Do More With Less"

More tokens isn't always better — curated information usually works better than dumping everything.

What Are Tokens?

Tokens are the smallest units AI processes. Rough estimates: - 1 English word ≈ 1.3 tokens - 1 Chinese character ≈ 2 tokens - 1 line of code ≈ 10-20 tokens

Claude's context window: 200K tokens GPT-4's context window: 128K tokens

Sounds like a lot? It fills up fast.

Why Manage Tokens?

More tokens = more distraction: When context is huge, AI may struggle to find relevant information.

Tokens = money: API calls are billed per token.

Tokens have limits: Exceed the context window and AI "forgets" earlier content.

Suggested Allocation

Experience-based suggestion (not a validated standard):

Task description:  10% (20K tokens)
Relevant code:     40% (80K tokens)
Examples/refs:     20% (40K tokens)
Constraints/rules: 10% (20K tokens)
AI output space:   20% (40K tokens)

Practical Tips

1. Summaries Instead of Full Files

❌ Give AI a 500-line file
✅ Give a 10-line summary + the 50 critical lines

2. Type Signatures Instead of Implementations

❌ Give the entire function implementation
✅ Give the signature and key logic

3. Diffs Instead of Full Files

❌ Give before and after versions of a file
✅ Give only the changed parts

4. Use CLAUDE.md for Persistent Context

# CLAUDE.md
## Project
React + TypeScript e-commerce app

## Stack
- React 18, TypeScript 5
- Zustand, Tailwind CSS

A few hundred tokens that persist across every conversation.

5. Process in Batches

❌ Ask AI to handle 10 files at once
✅ Split into 3 batches of 3-4 files

Tools

Common Mistakes

1. Too much code — "here's my entire project"
2. Too little context — "write a function" with no specs
3. Repeating yourself — re-describe project every conversation instead of using CLAUDE.md

Summary

1. Curate: Only give what AI needs
2. Structure: Use summaries, signatures, diffs
3. Persist: Use CLAUDE.md for project context

These are experience-based suggestions. Try different strategies and see what works best for your project.

Token management is a practical topic. If you have better tips, share them on GitHub.