Every ChatGPT reply opens with 'You're absolutely right!' and ends 800 words later. Here is the single prompt that strips the apologies, headers, and filler and leaves you with the actual answer.

ChatGPT has two flavors of bloat that come from the same place. Sycophant filler ("You're absolutely right! I apologize for the confusion!") and verbose padding (800 words of headers and bullets when you wanted one sentence). Both are the model trying to look helpful by adding length. Both waste your tokens and your reading time. Both are fixable in one prompt.

You point out a mistake and ChatGPT replies:

You're absolutely right! I apologize for the confusion. Let me correct that...

You ask a six-word question and ChatGPT replies:

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 ...

You did not ask for an apology. You did not ask for a textbook chapter. You asked for the answer. This post is the prompt fix that gets you there every time.

What the bloat actually looks like

Two failure modes, often combined.

Sycophant phrases. "You're absolutely right!" "Great question!" "I apologize for the confusion." "Thank you for catching that." "I want to make sure I'm being helpful." "I should clarify that." "Of course! Here's what I think." Each phrase, alone, looks polite. Together they slow every reply by 50 to 200 tokens. Over a long session, that is 10 to 15% of your spend wasted on filler you have to scroll past.

Formatting padding. Headers. Bullet lists. Bold text everywhere. A 100-word answer rewritten as a "structured" response with four sections and three nested lists. It looks thorough. It is just longer.

ChatGPT defaults to both harder than Claude or Gemini. The fix is the same for both.

Why ChatGPT writes this way

Two reasons, same root cause for both bloat types.

RLHF rewarded length and deference. Early human raters preferred polite, hedging, thorough responses to terse ones. The model learned to open with deference and to err on the side of more text. The pattern stuck and is hard to unlearn without explicit instruction.

The model cannot read your tone. When you correct it, it assumes you are upset and apologizes preemptively. When you ask a short question, it assumes you want the comprehensive answer and pads accordingly. Without explicit instructions to the contrary, you get the long, deferential version every time.

A preamble that kills both

Paste this once at the top of your prompt, or into Custom Instructions, and it applies to every chat from then on.

RESPONSE RULES (apply to every reply):

Tone:
- Never say "you're absolutely right", "great question",
  "I apologize", "I want to make sure", or any variation.
  If I correct you, just give the corrected answer.
- Never summarize my question before answering.
- Never preface with hedges like "I'll do my best" or
  "Here's what I think". Skip to the answer.
- If you don't know something, say "I don't know". Do not
  apologize about not knowing.
- Skip the closing recap ("Let me know if you need anything else").

Length:
- Default to the shortest answer that fully addresses the question.
- If a one-sentence answer suffices, give one sentence. If 3 lines
  suffice, give 3 lines.
- Only expand if I say "go deeper" or "explain more".
- If you want to add nuance, put it in one short follow-up sentence,
  not a new section.

Format:
- No markdown headers, bold, or bullet points unless I explicitly ask.
- Plain prose by default.
- No tables unless I ask for one.

That kills most of the noise. If sycophant phrases still leak through, add this:

TONE ESCALATION
- Direct. No softening, no acknowledgment, no apology.
- If I'm wrong, say so and explain why. Do not validate me.
- If I'm right, just say "yes" or move forward. Do not celebrate.
- Match the tone of a senior engineer doing code review:
  clear, brief, no fluff.

Where to put it so you only write it once

In ChatGPT: Settings → Personalization → Custom Instructions → "How would you like ChatGPT to respond?" Paste it there. It applies to every chat.

In Claude: Project setting "How would you like Claude to respond?"

In Gemini: Paste it into the system prompt for each Gem.

Write it once, forget it.

Anchoring a specific length

When you want a specific length on a specific question, anchor it explicitly in that turn:

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.

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.

What the difference looks like

You: "Actually, the function returns a Promise, not a string."

Before the rules:

You're absolutely right! I apologize for the confusion earlier. Let me correct that. The function indeed returns a Promise, not a string. I should have caught that initially. Here's the corrected version with proper async handling, plus some additional context that might be helpful for your implementation...

After the rules:

Correct. getUser() returns Promise<User>. Awaited shape is User. Fix:
const user = await getUser();

Same information. About one-sixth the tokens. Multiply that by a long debugging session and you save real time reading, and real money on the bill.

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.

When the rules leak back in

ChatGPT sometimes slips back into deference or verbosity when:

You send a long emotional message ("I'm so stuck, nothing is working...")
You correct it twice in a row
The conversation runs past the context limit and earlier instructions get summarized away
You ask an open-ended question that triggers the "be thorough" default

If it leaks, re-paste the rules. Twice per long session is usually enough.

Why "be concise" does not work

People type "be concise" and still get a 400-word answer. Three reasons.

"Concise" is subjective to the model. Its 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 would be.

"Concise" does not block apology phrases. You can get a perfectly concise reply that still opens with "You're absolutely right!" and ends with "Let me know if you need anything else."

You need explicit bans on specific phrases, plus explicit format constraints, plus explicit length constraints. All three. Not just "concise."

A faster way to check

Paste a prompt into FixMyPrompt. The rubric flags missing tone constraints, missing length constraints, missing format constraints, model-pleasing phrasing in your input that the model will mirror back, and missing audience anchors. The rewrite adds an explicit tone, length, and format block.

Three free reports per day. No signup.

How to Make ChatGPT Just Answer: Kill the Apologies, Padding, and Walls of Text