You typed: "What's the difference between TCP and UDP?"
ChatGPT gave you:
Great question! TCP and UDP are both fundamental transport-layer protocols in the OSI model, each serving distinct purposes in network communication. Let me walk you through the key differences:
1. Connection Establishment TCP is a connection-oriented protocol, meaning it establishes a connection between the sender and receiver before transmitting data...
And then eight hundred more words, with four headers and three nested lists.
You wanted one sentence. ChatGPT gave you a textbook chapter.
Verbosity is the default. Brevity requires fighting for it.
Why ChatGPT writes 600 words to a 6-word question
Two main reasons.
RLHF rewarded thoroughness. Early human raters preferred complete, structured answers. The model learned that longer reads as better.
The model has a formatting bias. Headers, bullet lists, and bold text get added because they look helpful, even when they make the answer four times longer than necessary.
Without explicit length and format constraints, you get the long version. Every time.
A brief-mode preamble
Paste this once into Custom Instructions or at the top of any chat:
RESPONSE RULES:
- Default to the shortest answer that fully addresses the question.
- No preamble. No "great question." No recap of what I asked.
- No headers, bold, or bullet points unless I explicitly ask for them.
- No closing recap or "let me know if you have more questions."
- If a one-sentence answer suffices, give one sentence. If 3 lines
suffice, give 3 lines.
- Only expand if I ask "go deeper" or "explain more."
- If you want to add nuance, put it in one short follow-up sentence,
not a new section.
That kills most of the bloat. Every chat starts with it from then on.
Anchoring a specific length
When you want a specific length, anchor it explicitly:
Answer in 1 sentence.
Answer in 30 words or fewer.
Maximum 50 words. Count silently before responding.
"Count silently before responding" sounds odd. It works. It forces the model to check its own length before generating, which catches the long-version impulse.
Killing the formatting bloat
Word count is not the only contributor to long responses. Headers and bullet lists turn a 100-word answer into a 400-word document.
Add this line:
Answer in plain prose. No markdown headers. No bold. No bullet points. No tables. Just sentences.
Output volume drops by 30 to 50% with no loss of information.
Telling the model who you are talking to
Length is bounded by audience. Different framings produce different defaults:
Explain it to a senior engineer who already knows the basics.
Explain it to a CFO who needs to make a yes/no decision.
Explain it like a Slack reply, not a doc.
The strongest version of this:
Explain it as if you're texting a coworker.
Texts do not have headers. Texts do not have bullet lists. Texts are one to three sentences. The model can match that tone. It just has to be asked.
When you actually want length
Sometimes you want the full version. The reverse instruction works too:
Explain this in full detail with examples. Do not truncate. I want the comprehensive version.
Or for a deep dive on a specific point:
Go deeper on point 2. 200 to 300 words. Include a worked example.
You can flip back and forth in the same chat. Brief by default, comprehensive on demand.
Why "be concise" does not work
People type "be concise" and still get a 400-word answer. Two reasons.
"Concise" is subjective. The model's idea of concise is "I cut the intro." Yours is "one sentence."
"Concise" does not override formatting. The model can write a "concise" answer that still has four bullet points and two headers, making it longer than plain prose.
You need a word count, plus a no-formatting instruction, plus a no-recap instruction. All three. Not just "concise."
A faster way to check
Paste a prompt into FixMyPrompt. The rubric flags missing length constraints, missing format constraints, open-ended questions where "go wide" is the model's default, and missing audience anchors. The rewrite adds explicit length, format, and audience constraints.
Three free reports per day. No signup.