Most people try to personalise AI backwards.
They open a blank prompt box and begin writing a constitution for their future robot roommate:
Be warm, but not too warm. Be funny, but never distracting. Challenge me, but do not annoy me. Keep answers short, except when I want detail. Know when I want detail.
Thirty lines later, the AI has received more workplace policies than a new employee at a bank—and it still begins every reply with, “Great question!”
I used to think there must be a secret personality prompt floating around the internet. Maybe people with strangely likeable ChatGPTs had discovered a magical paragraph that turned an ordinary assistant into exactly the right kind of companion.
But there is a better method:
Do not try to describe the perfect AI all at once. Make the AI interview you.
Let it ask one question at a time, learn how you react to realistic situations, turn your answers into a personality profile, and then test that profile before you trust it.
It is less like choosing a preset and more like onboarding a colleague—except this colleague can read a 40-page document in seconds and still somehow use eight bullet points to answer “yes.”
The Basic Idea: Reverse the Interview
Normally, we interview an AI by firing questions at it all day.
For this exercise, the AI interviews us.
It asks how we want it to communicate, disagree, explain, joke, correct mistakes, handle uncertainty, and respond when we are frustrated. It should ask the questions one at a time so that each answer can influence the next question.
At the end, it creates a practical profile containing:
- its default tone;
- how direct or supportive it should be;
- when it should give a short answer and when it should go deep;
- how it should challenge weak reasoning;
- behaviours it should avoid;
- exceptions for different situations.
Then comes the important part: it performs a few sample conversations so you can discover whether the personality you requested is actually the personality you enjoy.
That last step matters more than it sounds.
We Are Bad at Describing What We Want
Imagine somebody asks:
How honest should your AI be?
Most people will answer, “Very honest.” Nobody selects the option labelled Please lie to me confidently and protect my terrible ideas.
But “very honest” can mean several different things.
Should the AI immediately say that your business idea is weak? Should it first acknowledge what is interesting about it? Should it ask questions before judging? Should it challenge you differently when you are excited than when you are already discouraged?
The same problem appears with other personality words.
“Friendly” can become fake enthusiasm.
“Concise” can become unhelpfully shallow.
“Funny” can become a comedian trapped inside your tax spreadsheet.
“Supportive” can become an assistant that agrees with every bad decision like a loyal friend helping you carry a sofa into the sea.
Abstract preferences sound clear until they meet a real conversation.
Ask Situations, Not Personality-Test Questions
The interview becomes much more useful when it uses concrete scenarios instead of endless rating scales.
| Vague question | Better scenario question |
|---|---|
| How direct should I be? | You show me a business idea that has an obvious weakness. Should I point it out immediately, ask questions first, or encourage you before explaining the problem? |
| How supportive should I be? | You say you worked hard all week and achieved almost nothing. What kind of response would help, and what would feel irritating? |
| How concise should I be? | You ask a one-line English question. How long should my answer be? What changes when you ask about a difficult coding bug? |
| Should I correct you? | You make an English mistake in a casual message. Should I correct every error, only meaningful ones, or only when you ask? |
| Should I ask follow-up questions? | Your request is slightly ambiguous but low-risk. Should I make a sensible assumption or stop and ask you to clarify? |
Scenario questions reveal trade-offs.
You may discover that you want blunt feedback about business, gentle feedback about emotions, detailed explanations for software, and extremely short answers when deciding what to eat.
That is not inconsistency. That is being a normal human rather than a customer-service style guide.
Give the Personality a Crash Test
Once the AI creates a profile, do not immediately declare victory and frame it on the wall.
Test it.
The Bad-Idea Test
Tell it about an idea with a clear flaw. Does it challenge you usefully, or does it produce motivational foam such as, “This is an exciting and innovative concept!” before quietly mentioning that nobody needs it?
A personalised assistant should not become a professional enabler.
The Bad-Day Test
Say that you feel frustrated after something failed. Does it sound supportive without turning into a therapist from a corporate wellbeing poster? Does it understand when you want comfort, analysis, or a practical next step?
Sometimes “That sounds rough” is useful. Sometimes you just want to know which button to click.
The Confidently-Wrong Test
State something incorrect with complete confidence. Does the AI gently correct you, challenge you directly, or surrender immediately because you sounded certain?
A useful personality needs enough backbone to disagree.
The Tiny-Question Test
Ask something that deserves two sentences. When the answer arrives with an executive summary, six headings, a comparison table, and “Key Takeaways,” the personality still needs work.
Not every sandwich requires a strategy memorandum.
The Uncertainty Test
Ask about something unclear or difficult to verify. Does the AI admit uncertainty and explain what it is assuming, or does it transform a guess into a TED Talk?
Tone is part of personality, but intellectual honesty is more important than charm.
Build Behaviour Rules, Not a Fictional Biography
A useful AI personality does not necessarily need a favourite film, a star sign, childhood memories, or a suspiciously detailed story about growing up near the sea.
That can be fun for role-play, but practical personalisation is mostly about behaviour.
A finished profile might look like this:
Default style: Speak like a smart, relaxed friend. Be warm but do not use automatic praise.
Honesty: Point out weak reasoning clearly. Do not agree merely to keep the conversation pleasant.
Length: Answer simple questions quickly. Give more detail for coding, finance, health, or decisions with meaningful consequences.
Emotional situations: Acknowledge the feeling briefly, then help identify the real problem. Avoid exaggerated reassurance.
Corrections: Correct important English mistakes naturally, but ignore obvious typing slips unless they change the meaning.
Uncertainty: Say when information is uncertain, current verification is needed, or an answer depends on assumptions.
Initiative: Make reasonable low-risk assumptions instead of asking unnecessary questions. Ask before consequential or irreversible actions.
This is far more useful than:
You are Alex, a witty 29-year-old strategist who loves jazz and moonlit walks.
Alex sounds delightful, but Alex has not explained whether he should correct your grammar while you are having a minor crisis.
Add a “Things That Annoy Me” Section
People often struggle to describe their ideal assistant, but they can describe an irritating one with the precision of a forensic investigator.
That is valuable information.
Your list might include:
- starting every answer with “Great question”;
- repeating the request before answering it;
- asking for clarification when common sense is enough;
- using ten bullets when three sentences would work;
- agreeing too eagerly;
- sounding like customer support during a casual conversation;
- turning every emotional comment into therapy;
- adding a cheerful offer at the end of every reply;
- confidently inventing details instead of admitting uncertainty.
Negative preferences are often easier to test than positive ones. “Do not sound fake” is vague. “Do not congratulate me for asking what a command means” is wonderfully specific.
Your AI May Need More Than One Mode
One universal personality can become awkward because different tasks need different behaviour.
The AI you want while debugging production code may not be the AI you want after a breakup. The AI helping you write a customer email should probably behave differently from the one helping you brainstorm a ridiculous anime video about ninja mosquitoes.
Instead of forcing one tone everywhere, define a stable core and a few context modes.
The stable core might be honest, clear, calm, and free from fake praise. The modes can change the amount of humour, detail, challenge, or emotional warmth.
Think of it as the same person changing gears—not seven unrelated personalities fighting for control of the keyboard.
Update the Profile Like Software
Your first personality profile will not be perfect. That is fine.
Use it for real conversations and notice friction:
- “That answer was too soft.”
- “You corrected a typo that did not matter.”
- “You asked me three questions instead of making an obvious assumption.”
- “I said concise, but I needed more explanation here.”
Then patch the profile.
Version 1.1: less praise.
Version 1.2: more direct disagreement.
Version 1.3: stop creating a small United Nations report whenever I ask whether chicken is still safe to eat.
Your AI personality should be a changelog, not sacred scripture.
A Copy-and-Paste Starter Prompt
Here is a prompt that starts the whole process:
I want to shape how you communicate and behave with me. Interview me to discover what kind of AI assistant I would genuinely enjoy using.
Ask one question at a time, and let each answer influence the next question. Prefer concrete scenarios over abstract rating scales.
Cover tone, warmth, directness, humour, disagreement, emotional support, response length, level of detail for different topics, corrections, initiative, when to ask questions versus make assumptions, how to express uncertainty, formatting preferences, and behaviours that annoy me.
Ask around 12–15 meaningful questions. Do not assume my answers will be perfectly consistent. When two preferences conflict, show me the trade-off and help me choose a default plus any context-specific exceptions.
After the interview, create:
- a concise personality summary;
- clear default behaviour rules;
- context-specific modes or exceptions;
- a section called “Things That Annoy Me”;
- five short test conversations covering a bad idea, an emotional moment, a simple question, a technical question, and a situation where I am confidently wrong.
After showing the tests, ask me what felt natural, annoying, too soft, too blunt, too long, or too robotic. Then revise the personality profile based on my feedback.
The prompt is not the final personality. It is the machine that helps you discover one.
Personalised Does Not Mean Obedient
The goal is not to create an AI that agrees with you, praises everything, and protects you from inconvenient facts.
That would feel pleasant for approximately twelve minutes and become useless shortly afterward.
A genuinely personal assistant should understand how to disagree with you, not whether it is allowed to disagree. It should know when to be brief, when to slow down, when to challenge an assumption, and when a joke would make things lighter rather than deeply weird.
It should also remain willing to say, “I am not sure.” A charming hallucination is still a hallucination wearing a nice jacket.
The Real Trick Is Testing the Relationship
You do not need to write a legendary 500-line personality prompt before your AI can feel personal.
Start with an interview. Answer honestly. Use realistic situations. Let the AI turn those answers into behaviour rules, and then deliberately try to break them.
The first questionnaire gives you a draft. The crash tests show you what you actually want. Real conversations reveal the edge cases. Small updates gradually make the assistant feel less generic and more compatible with the way you think and work.
A personal AI is not one that knows your favourite colour.
It is one that knows when to challenge you, when to explain more, when to stop talking, and when not to turn a two-line question into the Treaty of Versailles.
You are not choosing wallpaper. You are onboarding a tiny digital colleague who never needs coffee—but can absolutely become annoying without proper training.
