In the last 17 days, I burned $6k worth of tokens on my Claude 20x and Codex 20x subs.

But I got so much work done by just fixing ONE issue.

I use Fable 5.1 (Extra High) for planning and GPT 6 Astra (High) for execution.

Both models think a lot and burn so many tokens.

I fixed this with the lean-thinking skill.

I forced Fable and Astra to just do this:

1. Understand the task first.
2. Make a proper execution plan.
3. Reuse what already works.
4. Build only what is needed.
5. Verify your work.
6. Fix any problems.
7. Stop when it is done.
8. Explain what’s built.

It saved me hours and almost 35% of tokens, with no quality degradation (I tested the same 11 builds with and without the skill, and the results were actually better).

I uploaded the skill to GitHub and am now open-sourcing it. Enjoy.

Here: https://github.com/Cjbuilds/lean-thinking…

Cjbuilds/lean-thinking

Source: https://github.com/Cjbuilds/lean-thinking

lean-thinking

Lean Thinking helps AI coding agents stay focused and finish the job. It tells them to plan enough, build what is needed, check the result, and stop.

Use it when an agent spends too long planning, keeps adding extra work, or rebuilds things that already work. Simple tasks stay simple. Risky tasks still get the checks they need.

How it works

• Understand the task first.
• Make a simple plan.
• Reuse what already works.
• Build only what is needed.
• Check that it works.
• Fix any problems found in the task.
• Stop when it is done and checked.
• Explain what changed and any remaining limits.

It follows your requirements and asks for permission when needed. It does not skip necessary work just to make the answer shorter.

Install

Clone this repository, then copy the skill folder into your project:

git clone https://github.com/Cjbuilds/lean-thinking.git

Replace /path/to/lean-thinking below with the clone location. Review an existing installed copy before replacing it.

For Codex:

mkdir -p .agents/skills
cp -R /path/to/lean-thinking/skills/lean-thinking .agents/skills/

For Claude Code:

mkdir -p .claude/skills
cp -R /path/to/lean-thinking/skills/lean-thinking .claude/skills/

Use

Invoke it by name in a request:

Use $lean-thinking to add CSV export to this report page.

The agent reads the relevant code, makes a plan, does the work, and checks the result.

Check the package

The checker uses only the Python standard library:

python3 scripts/check.py

It validates the skill metadata and package structure, required installation instructions, and local README links.

Files

lean-thinking/
├── assets/
│   └── banner.png
├── eval/
│   └── RESULTS.md
├── scripts/
│   └── check.py
├── skills/
│   └── lean-thinking/
│       └── SKILL.md
├── LICENSE
└── README.md

Limits

This skill cannot recover missing evidence, grant permissions, or make risky changes safe. Its result still depends on the agent’s tools, available context, and judgment. It does not promise to control private reasoning or reduce token use.

License

MIT. Use it, adapt it, and share it.