Skip to content
BSWEN
Blog

How to Reduce Token Usage in Codex CLI by 65%: The Caveman Approach

Cowrie
Cowrie
Dev @ Bswen

The Problem

I ran out of my weekly Codex CLI quota on Thursday. I had two days of work left but no tokens to spend. Every response burned precious quota:

Verbose Response Example
"I will now analyze the codebase structure and identify potential
optimization opportunities. First, let me examine the main entry
point and trace the execution flow through the various modules..."


[200+ tokens burned for one simple task]

The Reddit community confirmed this is a common problem:

Token Waste Pattern
- Verbose introductions: "I will now..."
- Explanatory padding: "Let me explain why..."
- Repetitive conclusions: "In summary, we have..."
- Over-detailed context: unnecessary background info

One user pointed out the real issue:

“We are all tokenmaxxing on tasks that don’t deserve it.”

Simple tasks don’t need paragraph responses. They need action.

The Solution: Caveman

Caveman is a Python tool by JuliusBrussee (48k+ GitHub stars). It forces AI responses into minimal “caveman english”:

Install caveman
pip install caveman

The result:

Before vs After Comparison
WITHOUT CAVEMAN (verbose):
  "I will now analyze the codebase structure and identify..."
  [~150 tokens]


WITH CAVEMAN (concise):
  "analyze codebase. find optimizations."
  [~50 tokens]


SAVED: ~100 tokens (65% reduction)

The key insight from Reddit:

“caveman…cuts ~65% of tokens by making responses caveman english”

“End-of-week when you’re scraping for quota this is the one”

How It Works

Caveman sits between your prompts and Codex. It compresses:

Compression Flow
Your prompt → Caveman filter → Codex receives minimal prompt
Codex response → Caveman filter → You receive concise response

The tool removes:

  • Polite phrases (“I will”, “Let me”)
  • Explanatory padding
  • Redundant context
  • Flowery language

Monitoring with ccusage

Caveman helps prevent waste, but ccusage catches runaway loops:

Install ccusage
pip install ccusage

Run it periodically:

Check usage
ccusage --today

The output shows:

ccusage Output
Session: 2026-04-28
Tasks: 12
Tokens used: 45,000 / 100,000
Remaining: 55,000 (55%)
Warning: Task #8 burned 8,000 tokens (tool loop detected)

A Reddit user explained why this matters:

“ccusage catches runaway tool loops that burn weekly limit”

Why This Matters

The math is simple:

Token Economics
Weekly quota: 100,000 tokens


WITHOUT optimization:
  - Verbose responses burn 150 tokens per task
  - ~660 tasks before quota exhausted
  - Runs out by Thursday


WITH 65% reduction:
  - Concise responses burn 50 tokens per task
  - ~2,000 tasks before quota exhausted
  - Lasts through Friday with buffer

That’s the difference between:

  • Quota exhausted Thursday afternoon
  • Quota lasting until Friday evening with margin
Working Time Impact
NO OPTIMIZATION: 4 working days per week
65% REDUCTION: 5 working days + weekend buffer


GAIN: 25-30% more productive time

Common Mistakes to Avoid

I’ve made all of these:

Mistake 1: Ignoring usage until quota runs out

Wrong vs Right
WRONG:
  Work all week → Thursday: "quota depleted" → Panic


RIGHT:
  Install caveman day 1 → Monitor with ccusage → Never hit quota wall

Mistake 2: Using caveman for complex reasoning tasks

Wrong vs Right
WRONG:
  Complex architecture decision → Force caveman style → Miss nuance


RIGHT:
  Simple task: "fix typo in config.py" → Use caveman → Save tokens
  Complex task: "design auth system" → Full response → Worth the tokens

Mistake 3: Not checking ccusage warnings

Wrong vs Right
WRONG:
  ccusage shows "tool loop detected" → Ignore → 8,000 tokens burned


RIGHT:
  ccusage shows warning → Investigate immediately → Stop runaway process

Mistake 4: Over-optimizing early in the week

Wrong vs Right
WRONG:
  Monday morning quota: 100% → Aggressive token saving → Mental overhead


RIGHT:
  Monitor usage daily → Activate caveman Thursday → Strategic optimization

When to Use Caveman

Decision Guide
[Is Thursday or Friday?]
  |-- YES → Activate caveman for all simple tasks
  |-- NO → Continue


[Is task complex reasoning?]
  |-- YES → Skip caveman, let Codex think fully
  |-- NO → Use caveman


[Quota below 30%?]
  |-- YES → Use caveman for everything except critical tasks
  |-- NO → Optional use

Summary

Caveman reduces Codex CLI token usage by ~65% through response compression. Combined with ccusage for monitoring, it prevents the Thursday quota exhaustion that plagues heavy users.

Key takeaways:

  1. Install early: Add caveman to your toolkit before you need it
  2. Use strategically: Activate Thursday/Friday or when quota drops below 30%
  3. Monitor with ccusage: Catch runaway tool loops before they burn your limit
  4. Match task complexity: Simple tasks → caveman; Complex reasoning → full responses

The Reddit thread that taught me this approach was blunt:

“End-of-week when you’re scraping for quota this is the one”

I’ve tested it. The 65% reduction claim is accurate. My quota now lasts through Friday with buffer for weekend tasks.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!

Comments

Tags: