2.5 Length is not clarity: keeping prompts tight

Why long prompts get worse

• More surface area for contradiction. By paragraph three you’re accidentally telling the model to be “detailed” and “concise” in the same prompt.
• Attention gets diluted. The mid-prompt lost-in-the-middle effect from 2.4 kicks in fast.
• Every rule adds latency. Big prompts are slower to send, slower to process, and more expensive per call.
• Debugging gets harder. A 30-line prompt that fails is hard to fix without changing three things at once.

The five things that pad prompts

1. Politeness rituals. “I would be so grateful if you could please…” adds no signal. The model treats them as filler.

2. Excessive framing. “You are a world-class expert with 40 years of experience…” rarely outperforms “You are a payments-support assistant.” Superlatives are noise.

3. Reassurances to the model. “Take your time. Think carefully. You can do this!” The model has no time and no morale. These lines cost tokens and change nothing.

4. Redundant negations. “Do not include preamble. Do not include an introduction. Do not start with ‘Sure!’ Do not add any pleasantries.” One line — “Return only the answer, with no preamble” — does the same work.

5. Repeated examples. Three near-identical few-shot examples that all show the same pattern. One representative example plus one edge case usually beats five clones.

The rewrite exercise

Before-and-after on a real prompt shape you’ll see in the wild:

Before (207 words):

Hello! I hope you're doing well. I have a very important task for you.

You are a super smart, world-class expert customer service assistant with
decades of experience handling e-commerce complaints for online shops in
South Asia. Please take your time and think very carefully.

I would like you to please categorize the customer message that I am
going to paste below. It is very important that you categorize it
correctly! Please pick the best fit.

The categories are:
- payment_failure (for payment problems)
- kyc_or_verification (for account verification problems)
- login_or_auth (for login problems)
- other (for anything else)

Please make sure to only return the category slug. Do not include any
preamble. Do not include any pleasantries. Do not include an
introduction. Do not add "Sure!" or anything like that. Just the slug.

Thanks so much!

Customer message:
"App keeps crashing when I try to send money."

After (61 words):

You are a support-triage assistant for a South Asian e-commerce platform.

Categorize the customer message into exactly one of:
- payment_failure
- kyc_or_verification
- login_or_auth
- other

Return only the category slug. No preamble.

<message>
App keeps crashing when I try to send money.
</message>

Same task, ~3× fewer tokens, more reliable output. Cheaper, faster, easier to debug.

The rule: every line should do work

Read your prompt line by line and ask: if I delete this, does the output actually get worse? If not, delete it.

When length IS the right answer

There are cases where a longer prompt genuinely helps:

• Many labelled few-shot examples for a nuanced classification task.
• Complex format specifications that need each rule spelled out.
• Domain glossaries the model wouldn’t know (specific product SKUs, in-house jargon).

The rule is not “shorter is always better.” It’s “every line should earn its place.” A 300-word prompt with 300 words of value is fine. A 300-word prompt with 60 words of value is not.

Check your understanding

Quick check

—

Which of the following lines can almost always be deleted from a prompt without harming the output?

1. A"Return the result as a JSON array with fields id (string) and score (number)."
2. B"Take your time and think carefully — you can do this!"
3. CA concrete before/after example demonstrating the exact output shape.
4. D"If a field is missing, return null. Do not invent values."

Check answer→