Prompt patterns that save usage.

Token efficiency isn't a virtue in itself — the goal is good work, not thrift. But the same habits that make you frugal with tokens also tend to produce sharper, faster answers. Efficient prompts are usually clearer prompts.

Ask specific questions instead of open-ended ones

"Tell me about X" burns tokens on preamble and background you didn't need. "What's the specific difference between X and Y, in three sentences" gets you the answer directly.

The model will match the shape of your question. If you ask an essay, you get an essay. If you ask for a specific structural output (a table, a bullet list, a number), you get that — and it uses fewer tokens.

Compress context before pasting

You don't need to paste the entire 40-page document to get a good answer about the third paragraph on page 12. Paste the relevant section, or if the document is short enough, paste the whole thing — but if it's genuinely long, do a first-pass summarization or extract just the pages you're asking about.

The single biggest usage waste is pasting a huge document and asking a small question about one line of it. The model reads all 40 pages every time.

Split long chats

A chat with 200 messages processes all 200 previous messages every time you send a new one. That's the linear growth pattern that eats usage.

If you're in a chat that's grown that long, start a fresh chat and paste only the relevant summary of what's come before. Alternatively, use a Project to hold the persistent context and let individual chats be short-lived.

Prune Projects periodically

A Project with 200 pages of attached context adds those 200 pages to every message. If half of that context is no longer relevant (last quarter's data, the old design spec, the deprecated API doc), remove it.

Once a month: Settings → Projects → open each → review attached files → remove what's no longer load-bearing. Takes 15 minutes and can cut a heavy user's usage by 30%.

Use system-style instructions once, not every message

If you always want responses in a specific voice or format, set it in the Project's Voice Guide once. Don't re-specify it in every message ("respond in bullet points... respond in bullet points..." — the tokens add up).

Use "stop and rewind" instead of re-asking

If a response is going in the wrong direction, interrupt it (button in the composer). Don't wait for the full 800-word response to load. Course-correct with "Actually, I meant X" — that's cheaper than re-sending the whole prompt from scratch.

The counterintuitive one: verbose prompts can save tokens

A long, specific prompt often produces a short, specific answer — because the model doesn't have to guess at what you meant. A short, vague prompt often produces a long, exploratory answer that hedges across possible interpretations.

Compare: - "Explain X" → 400-word essay that covers multiple angles. - "Explain X to a working backend engineer, focused only on the write-side consistency implications, in three paragraphs" → 200-word focused answer.

The second prompt is longer but produces a shorter response, and the total usage is lower.

What this looks like at scale

For a heavy Base user (~4 hours/day of active chat), the difference between prompt-efficient and prompt-wasteful patterns is often the difference between staying on Base indefinitely and needing Max 5x. That's $80/mo of savings from habit changes that also make your work better.