Caveman Prompting: How to Cut AI Response Tokens (With Examples)

Caveman prompting tells ChatGPT, Claude or Gemini to answer in compressed fragments, with no filler, preamble or recap, so responses use far fewer output tokens. Here is how it works, the three intensity levels, copy-ready instructions, and when not to use it.

By Quantara Editorial Team · Development · Published Sep 29, 2026 · 7 min read

Short answer: caveman prompting is a one-paragraph instruction, placed at the end of a prompt, that makes an AI model answer in short fragments with no pleasantries, preamble, hedging or recap. It keeps the facts and cuts the words. On typical explanatory questions Verbito estimates it removes about 60% of output tokens, with a lighter mode at about 30% and an extreme mode at about 75%. Code, numbers, names and quotes should always be exempt from compression.

Why AI answers waste tokens

Chat models are trained to be helpful and polite, which in practice means padding: "Great question!", a restatement of what you asked, a paragraph of context you already know, the actual answer, then a summary of the answer. Every one of those words is an output token. That costs you in three ways:

• Money: on API pricing, output tokens are usually priced at several times the input rate. GPT-6 Astra, for example, lists $10 per million input tokens and $50 per million output tokens (see our September 2026 pricing breakdown).
• Time: models generate output token by token, so a shorter answer arrives sooner.
• Attention: you have to read past the padding to find the part you needed.

What a caveman instruction looks like

This is the Caveman-level instruction Verbito's AI Token Saver appends to prompts. You can paste it at the end of any prompt yourself:

## Response style — Token Saver (Caveman)
Respond in caveman style: short fragments, answer first. Drop articles, filler words, pleasantries, hedging, and transitions. No preamble, no recap. Keep every fact, technical term, and step — cut words, not meaning. Never compress code, commands, numbers, names, quotes, or anything meant to be copied verbatim. If brevity would make the answer ambiguous or an instruction unsafe, spell that part out in full.

Two parts of it do most of the work. "Answer first, no preamble, no recap" removes the padding at both ends of the reply. The safety clause, "never compress code, commands, numbers…", stops the model from mangling anything you need to copy exactly, which is the main way naive "be brief" instructions go wrong.

Before and after

Question: when does Postgres choose an index scan over a sequential scan?

Default answer (165 tokens): "Great question! When it comes to database performance, there are a few important things to keep in mind. A sequential scan reads every row in the table, which can actually be quite efficient for small tables or when a large share of the rows match your filter. An index scan, on the other hand, uses the index to jump directly to the matching rows, which is usually faster when your filter is selective and only returns a small fraction of the table. In general, the query planner chooses between them based on its cost estimates, which depend on table statistics. If those statistics are out of date, the planner can make a poor choice, so running ANALYZE on the table is often a good first step. I hope this helps! Let me know if you'd like examples of how to read an EXPLAIN plan."

Caveman answer (57 tokens, 65% fewer): "Seq scan: reads every row. Cheap for small tables or when most rows match. Index scan: jumps to matching rows via index. Wins when filter selective (<~5–10% rows). Planner picks by estimated cost; stale stats → bad pick. Fix: ANALYZE."

Ultra answer (36 tokens, 78% fewer): "Seq scan = read all rows → small table / low selectivity. Index scan = jump via index → selective filter. Planner: cost estimate; stale stats → ANALYZE."

Token counts were measured with OpenAI's o200k tokenizer (the one GPT-4o and later models use). The answers are illustrative examples written to cover the same facts; savings on your own questions will vary by model and question.

All three answers carry the same facts, including the practical fix (run ANALYZE). What disappears is the padding: the compliment, the throat-clearing, the restated question and the closing offer. That padding is most of what you were paying for.

The three levels

The savings figures are estimates for typical explanatory answers. Real savings depend on the model and the question: they are largest on questions that normally trigger long, padded replies and smallest on answers that are already mostly code or data.

When not to use caveman prompting

• Customer-facing writing. Emails, landing pages, blog posts and ads need tone and full sentences. Use Lite or turn it off.
• Teaching beginners. Fragments assume the reader can fill in the gaps.
• Image and video generation. Midjourney, DALL-E and video models do not write text answers, so there is nothing to compress. Verbito disables Token Saver for these automatically.
• Safety-critical instructions. Medical, legal or operational steps must stay unambiguous. That is why the directive tells the model to expand whenever brevity would create ambiguity.

Does it cost input tokens?

Yes, a little. The Caveman instruction is roughly 100 input tokens. Input tokens are cheaper than output tokens and the saving on a single long answer is usually larger than that, so it pays for itself on any question with a substantial reply. For very short answers (a yes/no, a single number) it is not worth adding.

How Verbito scores token efficiency

Verbito's prompt score now has seven dimensions: clarity, context, specificity, constraints, output format, reusability and token efficiency. Token efficiency measures whether the prompt is lean and whether it bounds the length and format of the answer. A "tell me everything about X" prompt scores low; a prompt with an explicit length limit or a Token Saver directive scores high. Turning Token Saver on guarantees a token-efficiency score of at least 80 (Lite), 90 (Caveman) or 95 (Ultra).

Try it

Open the AI Token Saver, describe what you want, and pick Lite, Caveman or Ultra. Token Saver is also available in the full Prompt Generator and every Verbito tool page, free on every plan.