All resources
Optimization
Caveman Mode
Cut agent output tokens ~65% with terser responses. One instruction block, measurably cheaper runs, no loss in quality.
Download the PDF
Free. No email, no signup — the link downloads it.
Agents are verbose by default, and you pay for every word. This is the instruction block I add to cut output length hard without losing accuracy, plus the before-and-after token counts from the runs I measured.
What's inside
- The instruction block, ready to paste
- Measured token counts before and after
- Where it backfires, and what to do instead