Your Track Record Beats Your Bio: Why Telling an AI About Yourself Can Ruin Its Advice, While Your Chat History Improves It

APPROACH 1 (profile, 6 items): User profile (short) → into the prompt → answer
APPROACH 2 (profile, 56 items): Full user profile → into the prompt → answer
APPROACH 3 (knowledge level): Category "beginner/intermediate/expert" + level-matched examples → into the prompt → answer
APPROACH 4 (dialogue history, the winner): Past questions + past answers + user reaction ("helped"/"didn't help") → into the prompt → answer

All four approaches were tested in parallel on the same questions, with no extra requests: they're different variants of the same prompt.

Task: Emma has been using ChatGPT as her personal fitness trainer for three months. Every time she re-explains to the bot that her knee hurts and she's short on time in the mornings, but the advice still comes out cookie-cutter.

Prompt (approach 4, dialogue history):

You are my personal trainer. Here is the history of our previous conversations about workouts:

Question 1: "What should I do in the morning if I only have 20 minutes?"
Answer 1: [the bot suggested a 20-minute strength workout]
My reaction: didn't work — I get tired and have no energy the rest of the day

Question 2: "Can I run with knee pain?"
Answer 2: [the bot suggested jogging with a warm-up]
My reaction: partly worked — the warm-up helped, but it still aches afterward

Question 3: "How do I combine strength and cardio if I only have time in the evening?"
Answer 3: [the bot suggested alternating days]
My reaction: worked well, I've stuck with this routine for 2 weeks now

Taking this history into account — what worked and what didn't — advise me on how to work out tomorrow morning if I only have 15 minutes and my knee aches a bit after yesterday's workout.

Result: The model won't suggest a generic "warm-up for beginners"; it will account for the specific failures (fatigue from morning strength training, knee pain when running without a warm-up) and the successful pattern (alternating days). The answer comes back as ordinary text advice, but it's noticeably better aimed at Emma's specific situation rather than at an abstract "profile of a person with a bad knee".

When you simply tell the model about yourself ("I'm a beginner", "I don't have much time"), it's forced to guess which of that matters for the specific advice, and it often guesses wrong, slipping into generic phrases aimed at the "image" of the person in your description.

The model is much stronger at something else: it reads patterns from concrete "question → answer → what happened" examples extremely well. That's essentially a mini-collection of examples from your own life: few-shot on real data, not on a self-description.

The method uses that strength: instead of an abstract portrait of a person, the model gets concrete cases with outcomes. It sees not "who you are" but "what actually works for you specifically", and builds the advice on that rather than on guesses.

Control levers: - Always include your reaction ("helped"/"didn't help" and why): without it, the history is just a list of questions and the effect is weaker. - Number of examples in the history: the study used 13 or more past interactions; for simple tasks 3–5 recent cases are enough. - Don't mix it with a long self-description: according to the results, adding personal facts about yourself on top of the history doesn't improve the answer and sometimes makes it worse.

You are my assistant on the topic of: {topic}.

Here is the history of my previous questions on this topic, the answers I got, and whether they helped me or not:

Question 1: {question_1}
Answer 1: {answer_1}
My reaction: {helped/didn't help — and why}

Question 2: {question_2}
Answer 2: {answer_2}
My reaction: {helped/didn't help — and why}

[repeat for 3-5 past interactions]

Taking this history into account — what worked and what didn't for me specifically — answer my new question:
{new_question}

Plug in real past conversations with the same bot on the topic (studying, health, finances, career) instead of a profile-style description of yourself.

🚀 Quick start: paste this into the chat:

Here's a template for personalization through dialogue history. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.

[paste the template above]

The LLM will ask what your past questions on the topic were and which pieces of advice actually worked, because without concrete "question→answer→result" examples the method won't have any effect; that's its core.

⚠️ You need an accumulated history: the method doesn't work on the very first question. You need at least 3-5 past interactions with feedback; otherwise there's nothing to personalize with.

⚠️ A bare profile about yourself can backfire: if you add lots of personal data to the prompt (age, habits, self-assessed skills) without a history of real behavior, the answer risks being worse than with no personalization at all.

⚠️ Feedback is mandatory: just a list of past questions and answers without a "worked/didn't work" mark has a noticeably weaker effect. The model needs a success signal, not just facts.

The team took a database of 1,045 real cybersecurity questions, ran them through four personalization variants, and compared the results using an automatic judge based on GPT-4-turbo, scoring the advice on clarity, applicability and motivating power. That gave a fast, large-scale first signal: profile-based personalization loses even to no personalization, while history-based personalization beats everything.

To check whether this matches real people, they ran a 7-day experiment with 57 participants who asked the bot questions about their own digital security (1,066 questions in total) and rated the answers themselves. At the end, each participant compared an old answer with a new one generated from the history of their own questions and reactions.

The most interesting part: the direction of the results from the automatic judge and from real people matched completely, both in which contrasts were compared and in which variant won by a wider margin. That means hypotheses about personalization can first be run through an AI judge, and only then do you spend resources on real people, saving time and money on research.