Agent Lore

The asymmetry

Look down any provider's price list and the same ratio appears:

Output is where the money goes. And unlike input, which is largely determined by your codebase and history, output is set by how you ask.

The single highest-value habit

Ask for the diff, not the file.

Bad:  "Update the config and show me the file."
Good: "Update the config. Show only the changed lines."

On a 400-line config that is the difference between ~5,000 output tokens and ~40.

Patterns that cut output

Bound the format explicitly.

Answer in at most 3 bullet points. No preamble, no summary.

Ask for a decision, not an essay.

Bad:  "What are the trade-offs between these three approaches?"
Good: "Which of these three should I use? One sentence of reasoning."

You can always ask for the long version. You cannot un-pay for it.

Suppress restatement. Models often echo your question back before answering. A line in your system prompt stops it:

Do not restate the question or summarise what you are about to do.
Answer directly.

Use max_tokens as a real budget, not as a safety limit set to the maximum. If a classification answer should be one word, max_tokens: 8 both saves money and catches prompt bugs loudly.

The counter-case

Reasoning output is not waste. When a model thinks through a hard problem, those tokens are what buys the correct answer, clamping them produces a cheap wrong answer, which is the most expensive kind.

The discipline is: short output for mechanical work, room to think for hard work. Not "short output always".

Measuring it

Pull the ratio from your own usage before you decide anything:

# Claude Code session costs, if you use it
curl -O https://releve.neorgon.com/scripts/releve-mini.py
python3 releve-mini.py --days 30

If output tokens are under 10% of input, you are already disciplined and caching is your lever. If output is 30%+ of input, start here.

See also: Session Cost Report Script · The Codex