Home / Learn / Data & AI / Offline AI / Lesson 3.6
All lessons

Module 3: Tokens: the unit of everything · Lesson 3.6 · about 20 min

Saving tokens 1: short, clear prompts

The cheapest token is the one you never send or receive. How to write prompts that are short, clear and get short answers, without losing quality.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 3 · about 20 minutes · read first: 3.4. You’ll be able to cut a prompt’s tokens and its answer’s tokens without hurting the result.

The cheapest token is one you never send or receive, and the biggest saving is asking for a short, exact answer.

1. Why this matters

On a subscription, saved tokens mean you hit limits later. On an API, they are money. On your own machine, they are time: a job that writes half as many tokens finishes in about half the time.

2. Ask for less output

Output costs most (Lesson 3.4). Say exactly what shape you want, the way you would brief a busy colleague: “Just the total and the due date, please”:

  • “Summarise in 3 bullet points, at most 15 words each.”
  • “Reply with only the JSON, no explanation.”
  • “Answer yes or no, then one sentence why.”
  • Set a maximum tokens limit in the request as a safety net (Lesson 3.8).

3. Send less input

  • Cut the padding: “I was wondering if you could possibly help me to…” adds nothing; “Summarise:” works.
  • Say it once: standing instructions go in the system prompt, not every message.
  • Send only what is needed: the relevant section, not the whole document (Lesson 3.7).
  • Prefer plain text to pasted formatting, HTML or long tables when they are not needed.

4. Keep quality

Short is not vague. Keep what the model needs: the task, the audience, the format, and any rule it must follow. One example of the exact output format is often worth its tokens when the format is unusual.

Before (about 70 tokens) After (about 25 tokens)
“Hi! I hope you’re well. I have an email below from a customer and I’d really appreciate it if you could read it carefully and tell me what they want, and maybe how urgent it seems, thanks so much!” “Customer email below. Reply with: Request (one line), Urgency (low/medium/high).”
Maths and principle

Principle: time and cost scale with tokens, and output tokens weigh more.

Equation: cost ≈ input × price_in + output × price_out; local time ≈ input ÷ read speed + output ÷ write speed.

Analogy: a taxi with a cheap flag-fall and an expensive meter: the shortest route saves most on the meter.

Try it: a job writes 400 tokens at 15 per second. You cut the answer to 120 tokens. Time saved? (400 ÷ 15 − 120 ÷ 15 ≈ 26.7 − 8 ≈ 19 seconds per job.)

input, output
tokens sent and tokens written for one job
price_in, price_out
cloud prices per token; output is usually several times dearer
read speed, write speed
tokens per second on your own machine; writing is much slower
Physical meaning
Each word the model writes costs time locally and money in the cloud, and writing costs far more than reading. Trimming a long answer to the part you will actually use is the fastest way to make a job cheaper, without changing the model or the hardware.

“Cutting ‘please’ and ‘thank you’ is how you save tokens.” Polite words cost a handful of input tokens, which are read in a blink. The real cost is a long answer. One instruction such as “3 bullets, 15 words each” can save hundreds of output tokens per request. Keep your manners; trim the answer.

5. Worked example

Made-up numbers. A recruiter screens 300 CVs a week with a local model. Step through the change.

Old prompt: "Give a detailed assessment." Answers average 450 tokens.

Old weekly time. 300 × 450 = 135,000 tokens. At 12 tokens per second: 135,000 ÷ 12 = 11,250 seconds, about 3 hours 8 minutes of writing.

New prompt: "Fit: yes/maybe/no; 3 reasons, 10 words each." Answers average 60 tokens.

New weekly time. 300 × 60 = 18,000 tokens ÷ 12 = 1,500 seconds: 25 minutes. Same CVs, same model, same decisions.

6. Practice

  1. Rewrite one of your own prompts to half its length. Compare answers.
  2. Add an exact output format to a prompt that currently gets long answers.
  3. Using the box, how much time do 1,000 jobs save if each answer drops from 300 to 100 tokens at 20 tokens per second?
Answers
  1. If quality dropped, add back the one instruction that mattered, not the padding.
  2. Formats like “3 bullets, 15 words max” are followed well by most instruct models.
  3. 200 ÷ 20 = 10 s per job, 10,000 s in all: about 2 hours 47 minutes.

New words: output format, max tokens (a hard cap on answer length), padding (words that add tokens but no meaning).

  • Output tokens are the slow, expensive ones: ask for an exact, short shape (“3 bullets, 15 words each”, “only the JSON”).
  • Set a max tokens limit as a safety net.
  • Send less input: cut padding, put standing instructions in the system prompt once, send only the relevant part.
  • Short is not vague: keep the task, audience, format and rules; one example helps with unusual formats.
  • Time ≈ input ÷ read speed + output ÷ write speed, so halving the answer roughly halves the job.

Check yourself

1Which saves the most tokens on a typical job?
2You send the same long instructions with every request. What helps?
3When is a worked example in the prompt worth its tokens?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.