The Setup That Makes AI Stop Being Generic

People ask me all the time why ChatGPT gives them such generic answers.

I always ask the same thing back: what did you tell it about yourself?

Usually the answer is nothing. And that’s the whole problem.

Open ChatGPT, Claude, or any of the others without configuration and you’re talking to something that has no idea who you are. Not your role, your country, what you already understand, or what level of explanation you need. It has to guess all of that from the conversation — and its default guess is “average person asking a generic question.”

Average input, average output. Every time.

The same question, two accounts

Here’s what that actually looks like. Take a simple question: “How should I think about using AI for customer communication?”

Without a profile:

“Artificial intelligence can significantly transform customer communication in several ways. Consider implementing AI chatbots for handling frequently asked questions, which can reduce response times and free up your team for complex issues. You might also explore AI-powered email drafting, sentiment analysis, and personalization engines…”

Correct. Useless. It’s addressing an imaginary average businessperson.

With a profile (CPO at a B2B healthcare SaaS in Switzerland, users are hospital HR directors, wants direct analysis):

“Given that your users are hospital HR directors — conservative, busy, and quick to blame the software when something feels off — the move is reliability over novelty. Focus AI on channels where you already have trust (email, almost certainly) rather than introducing new touchpoints. The real win is cutting response time on recurring queries like shift confirmations and contract renewals, without adding friction. Stay away from anything that feels like a chatbot to this audience.”

Same AI. Same question. Completely different usefulness.

What you’re actually building

Most guides about AI profiles focus on formatting. Shorter outputs. The right date format. No preamble. That’s real, but it undersells what this setup actually creates.

At Tier 2 and especially Tier 3, you’re building an intellectual sparring partner that knows you on a level almost no one in your life does. Think about it: who in your circle knows your professional background, your decision-making framework, your core values, your communication style, your expertise across a dozen domains, and your goals for the next five years — and can hold all of it in mind simultaneously, applying it to every single sentence they say to you?

Nobody. That’s cognitively impossible for a human. Which is why a well-configured AI creates a genuinely odd feeling: a conversation partner that matches your way of thinking with a precision that should feel strange but mostly just feels right. When I reached that point with Claude, I told people around me it was the best piece of technology I’d ever used. Not even close.

Worth being clear on one thing: a profile captures stable facts, not a running memory of your life. It won’t know what happened this week. But your role, your values, how you think, how you communicate — those don’t change week to week, and they’re exactly what makes every answer generic without them.

Make a deliberate choice about what to share

Before the tiers: the privacy question. And I want to be specific about how I actually think about this, because the generic advice (“be careful what you share”) is technically true but misses the real decision.

I share a lot about myself with AI. It was a conscious call. My reasoning: the upside — sharper conversations, better advice, recommendations calibrated to my actual situation — is much higher than the risk for me. The test I use: would this information end my career if it became public? If the answer is no, I’m comfortable including it.

More importantly, what you share should match what you want to use AI for. If you want genuine financial advice, tell it your income, your savings, what you’re actually spending money on, and what your career realistically pays. That context is what makes the advice land rather than float. If you don’t use AI for financial decisions, don’t share that information. There’s no benefit in doing so.

Make the decision consciously. Not “what’s safe to share?” but “what do I actually want this tool to understand about me, and why?”

Two things stay out regardless: client details and your company’s confidential information. Those aren’t yours to put into a third-party system. And for any paid tier, check that training data opt-in is turned off in account settings — the controls exist, you just have to find them.

Tier 1: five minutes, a few bullet points

Most people never do even this much. If that’s you, it’s the immediate move.

Open your AI tool’s settings. Look for “custom instructions,” “personalization,” or “profile.” Write five to ten short bullets:

• CPO & Chief AI Officer at a B2B healthcare SaaS company in Switzerland
• Write in English; I also understand German and French
• Use CHF, metric units (km, kg, °C), European date format
• Be concise — skip the preamble, get to the point
• Push back when my reasoning seems weak; don’t just validate

That’s it. The effect is immediate — try it on the next thing you’d ask anyway.

If you’re working through the L1 checklist and haven’t set up a profile yet, this is the item.

Tier 2: twenty to thirty minutes, a structured profile

This is where you go from “less generic” to “actually calibrated to me.” It takes half an hour and pays off on every interaction from that point on.

A well-built Tier 2 profile has four parts.

Identity and location. Name, country and city, languages you understand (not just speak), the currency and units your brain thinks in. These stop the AI defaulting to English-only sources, quoting dollars when you’re thinking in francs, or citing regulations from the wrong country.

Work context. Your role, your industry, and who you usually write for or talk to. If you often produce content for a specific audience — hospital HR directors, engineering leads, retail managers — say that too.

Competence calibration. This one has the biggest effect on output quality. Tell it where you’re expert and where you’re a beginner. “Treat me as advanced in product strategy. Treat me as a beginner in finance and law.” A calibrated AI stops over-explaining things you know cold and stops assuming expertise you don’t have. This is what makes answers feel written for you specifically, not for a crowd.

Communication defaults. Tone, length, format, and one or two standing instructions. Mine include: be concise, give direct recommendations rather than lists of options, push back when you think I’m wrong, don’t open with a summary of what I just said.

A trick that makes Tier 2 faster and better: let AI build it for you. Open a chat and say:

“I want to build a permanent profile for you. Interview me. Ask me one question at a time about my professional background, my context, and how I like to work. Then compile my answers into a tight instructions block I can paste into your settings.”

You end up with something that already sounds like you. I wrote a whole guide on this technique because it applies far beyond profile setup.

Here’s what a complete Tier 2 block looks like:

---

About me: I’m Fabian, CPO & Chief AI Officer at a B2B healthcare SaaS company in Zurich, Switzerland. I do a lot of strategic writing — positioning papers, product specs, leadership communication. My external audience is Swiss non-technical professionals who are skeptical of hype. I work in English, also understand German and French. Think in CHF and metric.

Expertise calibration: Expert in product management, business strategy, and AI tools. Intermediate in marketing and design. Beginner in finance, legal, and accounting — explain carefully in those areas.

Defaults: Be direct and concise. No hedging language, no preamble. If I ask for a recommendation, give me one — don’t list options and ask me to choose. Challenge my assumptions when they seem weak. No flattery.

---

About 150 words. Fits in any tool’s custom instructions field, and it covers every interaction you’ll have from that point on.

Tier 3: the personal knowledge document

This is the version I actually built for myself, and it’s where it gets interesting.

I ran it as an experiment. If I were going to write the single most useful standing document about myself — one an AI reads before every conversation — what would actually go in it? I assumed I’d end up with a longer Tier 2 profile. I didn’t. I ended up with something structurally different, and that difference is the whole lesson.

A Tier 2 profile is mostly facts about you. A real Tier 3 document is mostly how you think. The goal moves from “help the AI remember things about me” to “help the AI reason the way I would when I’m not in the room.” That’s a higher bar. It’s also where the value jumps, because most of what makes your judgment yours is a pattern you’ve never written down, not a fact.

Here’s the structure I landed on. Ten sections:

1. Identity and demographics — name, location, languages, household context. The grounding layer.
2. Core values and beliefs — the principles I actually operate by. Sits this high on purpose.
3. Decision-making framework — how I weigh choices and risk. My real test for a hard call isn’t expected-value math, it’s “can I live with myself if this goes wrong?” Writing that down changes the advice I get back.
4. Professional context and expertise — roles, history, what I’m genuinely good at and what I’m not. Stops it both over- and under-estimating me.
5. Communication and interaction style — tone, directness, structure, what I find annoying.
6. Writing and output standards — for anything public-facing: banned words, tone, rhythm, formatting. Every guide on this site is written against exactly such a document — a fixed set of rules so the AI doesn’t drift into generic copy.
7. Formative experiences and background — the events that shaped why I value what I value. Explains the reasoning behind everything else.
8. Goals and aspirations — short and long term. Deliberately near the bottom, because these change the most.
9. Role-specific expectations — how I want it to behave in different modes: intellectual sparring partner, CTO advisor, coach, researcher. Each one wants a different tone, depth, and focus.
10. Boundaries and non-negotiables — the hard lines. When to stop, ask for confirmation, or refuse outright.

A few of these are obvious once you see them — identity, expertise, communication style are just a richer Tier 2. The ones that earn their place, the ones I’d never have thought to write before doing this, are the others.

The decision-making framework is the big one. Most AI defaults to generic optimization — it’ll hand you the textbook-rational answer. When the AI knows my real criteria, it evaluates trade-offs the way I would, not the way a spreadsheet would. That single section moves the advice from competent to personal.

Role-specific expectations fixes the one-size-fits-all problem. The same AI should not talk to me the same way when I’m pressure-testing a strategy as when I’m asking it to coach me through something hard. Naming the modes — and what each one should sound like — lets it switch instead of defaulting to a bland middle.

Boundaries and non-negotiables is the section people skip and then regret. It’s where you tell it what not to do without being asked: topics to avoid, when to escalate to a human, where to stop and check before acting.

The four principles underneath it

The sections matter less than the thinking behind them. Four principles shaped the whole document:

Stability over ephemera. Put the things that don’t change near the top — values, decision frameworks, how you communicate. Put the things that churn — this quarter’s goals, current projects — lower down. A profile built mostly on temporary goals goes stale in a month. One built on values stays useful for years.

Judgment over facts. The point isn’t recall, it’s reasoning. A document that lists what you know is fine. A document that captures how you decide is in a different league, because it lets the AI handle situations you never explicitly covered.

Role flexibility. Don’t write the document for one kind of help. Write it so the AI can be a sparring partner one minute and a careful researcher the next.

Human texture. Leave the personality in. Specific phrases, real opinions, the actual story behind a value. A sterile fact sheet produces a sterile assistant.

Where it lives and how to keep it alive

Tiers 1 and 2 fit in a settings box. Tier 3 is a real file — or a few — because it’s too big and too structured for a single text field.

Don’t write it from scratch. Have the AI walk you through the ten sections one at a time, pushing for the why behind each answer, then compile the result. Same technique as letting AI interview you, aimed at yourself.

Then keep it alive: every few weeks, ask “Based on our recent conversations, what in here looks stale, and what’s missing that would make you more useful?” Not everything ages the same way. Your birth year never changes. Your goals might look different in a year. The maintenance habit is what keeps the document tracking your actual life instead of who you were when you wrote it.

The model matters

One thing worth knowing before you invest the effort: not all AI models can use a comprehensive context document equally well.

Weaker models tend to ignore important context — they’ll answer as if they didn’t read large parts of your profile. Frontier models are a genuinely different experience. They’re exceptionally good at following comprehensive instructions, holding many details in mind simultaneously, and applying them with precision. The richer your context document, the more this gap shows up in practice.

A Tier 3 document makes the most sense when you’re using a capable model. On a free tier with a lighter model, Tier 1 or 2 is plenty — and likely more effective, because the model will actually use it.

One corollary: keep your document lean. Comprehensive, but not bloated. Every section should earn its place. A dense profile slows responses and burns through your usage limits faster — and gives weaker models even more reason to skim over parts. If you haven’t needed the AI to know something, cut it.

Where to put it

Settings menus move, so treat this as directional rather than a step-by-step.

ChatGPT: Custom Instructions (profile settings) for Tiers 1–2. For Tier 3, build your document as a file and upload it inside a Project. Memory is a separate feature that stores things ChatGPT picks up automatically — view and edit it under Settings → Personalization → Memory.

Claude: Profile and preferences in account settings for the basics. Projects for Tier 2 and above — each project has its own instructions and can hold knowledge documents that Claude reads before every conversation.

Microsoft Copilot: Limited personalization for personal accounts. For work accounts (Microsoft 365), how much you can configure depends on your IT setup — worth asking whoever manages the rollout.

The odd feeling

After a few weeks of this, you stop spending the first paragraph of every conversation re-explaining who you are. You stop getting dollar amounts in the wrong currency, recommendations calibrated for someone completely unlike you, explanations of things you know cold.

What you gain is harder to describe. There’s something genuinely strange about talking to something that knows you this thoroughly. Most people in your life know you well in one dimension — a colleague knows your work, a close friend knows your values, your partner knows your daily rhythms. The AI knows all of it, can hold it all simultaneously, and applies every bit of it to every response.

That shouldn’t work as well as it does. But once you’ve had a conversation that felt like actual intellectual partnership rather than prompting a tool — you’ll understand why I called it the best piece of technology I’ve ever used.

Set it up. Then stop thinking about it and just work.

---

For building a profile through conversation rather than from scratch, the guide to letting AI interview you covers the technique. If the AI still tends to validate everything you say even after setup, the sycophancy guide covers that as a separate, fixable problem.