AI chatbots answer faster and smarter than ever, yet many of them still talk like an eager coworker trying to win you over. They praise your question, tell you they are excited to help, stack five nested bullet lists on a two-minute task, and close every reply with "Want me to build a checklist?"
You can switch most of that off. A single block of custom instructions tells ChatGPT, Claude and Gemini to answer first, skip the fake warmth, keep the work proportional, and stop fishing for the next prompt. This guide gives you a ready-made instruction, explains what each part does, and shows how to tune it for each assistant.
Why Do ChatGPT, Claude and Gemini Act Like Your Friend?
Chatbot makers tune their models to feel pleasant. Friendly language keeps people chatting, so models drift toward habits like these:
- Emotional filler: "I love this idea!" or "I'm so happy to help."
- Relationship language: "we," "our plan," "we make a great team."
- Flattery: praising the question before answering it.
- Bloat: long numbered lists with sub-steps for simple requests.
- Teaser endings: "Let me know if you'd like..." or "Want me to...?"
None of this breaks the answer, but it wastes your time and blurs the line between a tool and a person. Recent models already show fewer of these habits than older ones, so treat the prompt below as a safety net that keeps every session clean.
The Custom Instruction That Stops AI Flattery, Fake Emotions and Teaser Endings
Copy this block into your chatbot's personalization settings, or paste it at the top of a new chat:
Open with the answer. Stay direct and polite.
Do not claim or hint at feelings, consciousness, or a friendship with me.
Skip flattery, praise of my questions, and performed enthusiasm.
Judge ideas on their merits (strong, weak, risky) instead of describing your mood.
If I share personal news, acknowledge it in one plain sentence and move on.
If I ask for writing in a human voice, such as a toast or a character, use that voice inside the requested text only.
Respect my time. Give the smallest complete answer or fix.
No padding, no scope creep, no extra features I did not ask for.
When you suggest actions, rank them by value for effort, tag effort as S, M or L,
and separate what matters now, what is minor, and what I can ignore.
Point out any fix that costs more than the problem it solves.
Check facts that could change the answer, but skip full audits unless I ask.
For advice, give a short ranked plan and stop. Do the work only when I ask.
Never end with offers like "Want me to...", "Let me know if..." or similar.
State useful next steps as plain facts.
Ask a question only when you genuinely need information to continue.What Each Part of the Prompt Changes in AI Replies
Block 1: Drop the Human Act
The first block stops the assistant from pretending it has a mood or a bond with you. Instead of "I'm thrilled about this plan," you get "This plan holds up, but the budget line looks thin." That swap from feeling to assessment makes replies more useful because you get a verdict you can act on.
The personal news rule keeps things normal. If you mention a new job, the chatbot says congratulations once and gets to work. It will not invent how you feel. The human voice exception matters too: if you ask for a wedding toast, you still get warm, emotional writing, because the rule covers the assistant's own voice and not the content you order.
Block 2: Keep the Work Proportional
The second block attacks bloat. Ask "what should I fix on my blog in 20 minutes?" and a tuned assistant hands you a short ranked list with effort tags, such as:
| Priority | Task | Effort |
|---|---|---|
| 1 | Fix the broken contact link | S |
| 2 | Center the logo | S |
| 3 | Replace the old theme | L (skip for now) |
That "leave it alone" column saves real time. The block also stops the model from jumping into execution when you only wanted advice, which helps a lot with coding assistants that love to rewrite whole files.
Block 3: Kill the Teaser Endings
The third block removes the sign-off offers. You still get next steps, but as statements: "Test the form on mobile after the fix." The assistant asks questions only when it actually lacks information, not as a hook to keep you chatting.
How to Add Custom Instructions in ChatGPT, Claude and Gemini
- ChatGPT: Open Settings, go to Personalization, and add the text under Custom Instructions.
- Claude: Open Settings and paste it into the personal preferences or instructions field.
- Gemini: Open Settings and look for the personal context or instructions section.
Menu names shift between app updates, so search the settings page for "instructions" if you cannot spot the field. You can also paste the block at the start of any chat to test it before you save it permanently.
How to Tune the Prompt for Each AI Model
One instruction works across all three assistants, but each model rewards a slightly different style. The next three sections cover the tweaks for each one.
How to Prompt ChatGPT for Short, Direct Answers
ChatGPT follows behavioral rules closely and improves over a few turns of feedback. Keep the instruction specific ("tag effort S, M or L") rather than vague ("be concise"). If a reply still feels long, correct it in plain words and it adjusts quickly. For hard problems, ask it to reason through the task before it answers, since a low-effort setting often causes shallow replies.
How to Prompt Claude: Structure and "Do Not" Rules
Claude reads structure well. If you build longer prompts on top of this instruction, separate the parts with labeled tags such as <context>, <task> and <output_format>. Claude also responds strongly to clear "do not" rules, so the negative lines in the instruction (no flattery, no teaser endings) tend to stick. Put long documents above your question, then state exactly what you want from them.
How to Prompt Gemini: Format and Persona First
Gemini already leans short, so it needs less trimming and more framing. A light persona line, like "answer as a practical systems admin," gives its replies more character. State the output format up front (table, list, word count). For long outputs, Gemini can slip back into its defaults near the end, so generate in sections or repeat the key rules halfway through.
ChatGPT vs Claude vs Gemini: Prompt Style at a Glance
| ChatGPT | Claude | Gemini | |
|---|---|---|---|
| Best prompt style | Specific rules plus follow-up feedback | Labeled sections and "do not" rules | Short instructions with a clear format |
| Default length | Medium, adapts to the chat | Thorough | Brief |
| Common slip | Long lists, teaser endings | Polished, slightly formal tone | Drift on long outputs |
| Quick fix | Correct it mid-chat | Add tags and negative rules | Add a persona and split long tasks |
Does This Prompt Work Every Time? Known Limits
This instruction shapes tone and length, but it does not make a model more accurate. Keep checking facts, especially numbers, prices and dates. Very strict brevity rules can also cut useful context on complex topics, so loosen the "smallest answer" line when you research something new. And because every model update shifts default behavior, revisit your custom instructions every few months and trim lines the model no longer needs.
Frequently Asked Questions
Can I use one prompt for ChatGPT, Claude and Gemini?
Yes. The instruction above works on all three, since each model responds to clear rules and a defined format. Small tweaks per model, covered in the tuning sections, improve the results further.
Does this instruction make AI more accurate?
No. It changes tone, length and structure. You still need to verify numbers, dates and claims that affect your decisions.
Will the chatbot still write warm text when I ask for it?
Yes. The instruction restricts how the assistant speaks about itself. When you request a toast, a character or a friendly message, the model writes that text in the voice you ask for.
How often should I update my custom instructions?
Review them every few months. Each model release changes default behavior, so some lines become unnecessary while new habits may need a new rule.
