Skip to the page
Kalibr For teamsFor couples Free 3-minute test Sign in For teamsFor couples
All posts June 2026 6 min read

Your AI doesn't know you. That's the whole problem.

AI advice is generic because the model has no information about you, not because it's agreeable. Those are different problems with different fixes.

AI advice is generic. The usual explanation is sycophancy: RLHF training rewards responses that feel good, so models learn to validate rather than challenge.

That's true. It's also the smaller problem.

A model that knew you well could still be sycophantic. The deeper issue is that it has no information about you. It uses a median-human prior. Answers calibrated for the average person. Which means they're wrong for any specific person in proportion to how different they are from average.

What the model is missing

When you ask Claude to review your business plan, it doesn't know you have a documented pattern of over-weighting revenue potential against operational complexity. It doesn't know you go quiet when feedback lands bluntly instead of pushing back. It doesn't know your most confident moments are historically your worst decisions.

Without that, it answers for someone it invented. The advice is for a fictional version of you.

Making models less agreeable on average doesn't fix this. It shifts the median. You're not the median.

An AI constitution

A constitution is a system prompt loaded before the conversation starts. It tells the model your reasoning tendencies, your documented failure modes, how you take feedback, and where you need challenge instead of confirmation.

With that context, when you exhibit a known pattern (treating a planning problem as a research problem because you're in your over-analysis zone), the constitution primes the model to flag it. When you're in your blind-spot territory, the model probes instead of agrees.

There's also a bypass. Type "free mode:" and the constitution switches off for that session. It's a prior, not a constraint.

How Kalibr generates yours

Twenty minutes. Four published instruments. Scoring is deterministic. The output is plain text you paste into whatever you use: ChatGPT, Claude, Gemini, anything with a system prompt field.

You own it. No ongoing subscription to keep using it.

Early users describe the shift as going from talking to a very smart stranger to talking to someone who already knew how they thought.

When it matters

The constitution adds less on well-defined tasks: coding, summarising, formatting. It adds a lot on open-ended, judgement-heavy work: evaluating a business decision, thinking through a conflict, finding where a strategy is weak.

If you use AI to think, rather than to execute tasks, the constitution changes the quality of that thinking.

Take the assessment

20 minutes. One assessment. A leadership profile, a blind-spot map, and an AI constitution built from your results.

See pricing