SOP: Cut AI Token Use with a Lean Prompting Method

Purpose and boundary

This SOP uses “Cave Man” as a shorthand for deliberately plain, short AI instructions. It is not presented here as a standardized or independently validated method. The operating idea is simple: remove words, context, and output requirements that do not change the result you need.

Token savings depend on the model, the conversation history, attached materials, tools, and the requested answer. This process does not guarantee a specific reduction.

1. Define the smallest useful outcome

Before drafting a prompt, write down:

• the task to perform;
• the input the model actually needs;
• the required output format;
• the acceptance criteria.

A narrow target prevents broad prompts that invite unnecessary explanation.

Verbose request:

Please act as an experienced editor and carefully review the following text. I would like you to identify any issues and provide a detailed response that I can use with my team.

Lean request:

Edit this text for spelling and clarity. Return: 1) corrected text, 2) a bullet list of substantive changes.

The second version makes the task and response shape explicit without adding a role description or undefined expectations.

2. Keep only decision-relevant context

Include context when it changes the answer. Remove it when it is merely background.

Use this test for each paragraph, document, or instruction: Would the result materially change if this were omitted? If not, leave it out.

When working from a long source, send the relevant excerpt rather than the whole document where possible. If the source must remain available, state what section should be used.

3. Replace prose with compact constraints

Convert repeated or conversational instructions into short constraints. A compact prompt can still be precise.

For example:

Audience: current customers. Tone: direct. Maximum: 120 words. Include one next step. Do not invent product details.

This is usually easier to inspect than a paragraph that repeats the same requirements in several ways.

Avoid compressing away critical safeguards. Constraints about accuracy, approved sources, privacy, legal review, or exact formatting should remain when they apply.

4. Set an output budget

Ask for the smallest response that can be used. Specify a word limit, a number of bullets, a table schema, or a fixed JSON structure when appropriate.

Examples:

• “Reply with three bullets.”
• “Maximum 80 words.”
• “Return CSV with columns: issue, priority, owner.”
• “If information is missing, ask one clarifying question.”

An output budget matters because generated text is also part of token use. It also makes review faster.

5. Avoid repeating stable instructions

In a multi-turn workflow, do not paste the same large brief into every new request unless it is necessary for the model to follow it. Maintain one concise source of truth for stable rules, then add only the task-specific change or input.

Before reusing a prompt, remove:

• duplicated objectives;
• repeated style rules;
• examples that no longer guide the task;
• prior outputs that are not needed for the next decision.

Do not remove information the model cannot otherwise access in the current interaction.

6. Use staged requests for complex work

A single all-in-one instruction can produce a long answer and make errors harder to isolate. Split a complex task when the intermediate result can be checked.

A practical sequence is:

1. Extract the facts from the supplied material.
2. Confirm or correct the extracted facts.
3. Produce the requested deliverable using only those facts.

This may not always reduce total tokens: additional turns add their own input and output. Use stages when control and review outweigh that cost.

7. Review the prompt before sending

Use this short checklist:

• Is the task stated in one sentence?
• Is every piece of context necessary?
• Are the output format and limit explicit?
• Have duplicated instructions been removed?
• Are critical accuracy or compliance constraints still included?
• Would a smaller input or narrower output still work?

Example template

Task: [single action]
Input: [only relevant material]
Output: [format and maximum length]
Constraints: [only rules that affect the result]
Missing information: [ask one question / state the gap]

The “Cave Man” approach is useful as an editing pass, not as a reason to make instructions ambiguous. Keep the minimum wording that preserves the required result, safeguards, and reviewability.