How to stop AI giving you generic answers about your own life
AI gives you generic answers because it is starting from zero. The fix is not a better prompt — it is two files that every AI session reads before it says anything.
Every time you open a new chat, the AI knows nothing about you. It gives the same answer it would give anyone. You re-explain your situation, it produces something reasonable, and you move on. The problem is not the answer. It is the cost of getting there — re-establishing context that should already be there, every single session.
There are two documents that fix this. The first tells the AI who you are. The second tells it how it must behave. Together they mean the AI starts every session already knowing what matters and already constrained from doing things you would not want.
Why re-explaining context every time is a trap
The mental model most people use with AI is: I describe my situation, the AI responds, I get a result. This works fine for one-off questions. It breaks down when you are using AI across your whole life — your finances, your health, your work, your kids, the project you are building in the hour after they go to bed.
For recurring situations, the context is the same every time. You are the same person with the same priorities and the same constraints. The AI just does not know that yet.
What you want instead is a setup where the AI reads your context before it says a word. Where the session starts oriented rather than blank. Where you do not have to spend the first five minutes explaining what you explained last week.
The about-me file: context that travels with you
The first document is an about-me file. It is a plain text file that lives in your life system and gets read at the start of every AI session you point at that system.
It covers a small number of things: who you are, what you are working on right now, what your biggest constraint is, and what is off the table — the things you are not doing this year and do not want to be asked about. It gets updated when something changes. It does not need to be long. It needs to be current.
The difference between an AI session with this file and one without it is the difference between talking to someone who already knows your situation and talking to a stranger. The AI asking a second question that shows it registered the first. The AI not suggesting something you have already ruled out. Small things, but they compound across hundreds of sessions.
The constitution: rules it cannot break
The second document is different in kind. The about-me file gives the AI context. The constitution gives it constraints.
A constitution is a set of rules that govern how every AI and agent working in your system must behave. Not a personality setting. Not a system prompt. A document that gets read on startup and defines the limits of what the AI is allowed to do.
Some rules in a constitution are about communication — how you want information delivered, what to skip, whether to lead with the conclusion or build up to it. Those rules matter but they are not the ones that protect you.
The rules that protect you are about action. What the AI is not allowed to do without asking first. What it is not allowed to put into a chat under any circumstances — bank details, passwords, your children's names. What it must never send, post, or publish without explicit approval. What it should do instead of deciding when it hits a limit.
The more detailed you go with the answers, the more detailed the constitution.
Chapter 5.4 · 3:47
That is also the caution. A constitution built from vague principles is not much protection. The rules that matter are the specific ones — the ones written because you have thought through the failure mode they prevent.
Why the constitution matters most when agents are running
If you are using AI for occasional questions, a constitution is useful but not urgent. The moment you start running agents — tasks that execute while you are not watching — it becomes essential.
Agents follow instructions. They do not make moral judgements. If an agent is researching on the internet, reading emails, or interacting with external services, it will encounter content written by people who want it to do something other than what you asked. This is called a prompt injection: a set of instructions embedded in something the agent reads, designed to redirect its behaviour.
A well-written constitution does not make agents immune to this, but it limits the damage. An agent that cannot send anything without approval cannot be tricked into sending anything without approval. The rule holds because the rule is written down.
What the constitution is not
A constitution is not a prompt that makes the AI perform a persona. It is not the list of things you want the AI to do for you. It is not a wish list.
It is also not something you write once and assume is working. The gaps in a constitution reveal themselves when an agent does something you would not have wanted. The rule you forgot to write. The permission you assumed it did not have. Each of those is a rule to add.
Writing a constitution is partly about knowing yourself well enough to know what you need to protect. Some of that knowledge only comes from using the system long enough to see what goes wrong.
The interview that builds both documents
The approach taught in the course uses an AI interview to generate both documents in a single session. The AI asks a small number of questions about your priorities, your constraints, and how you want to be communicated with. Your answers become the about-me file. The rules implied by those answers become the first draft of the constitution.
This is faster than writing either from scratch, and it surfaces things you would not have thought to include. When you say your biggest constraint is money, the AI might flag that your money folder should be kept out of sessions that do not need it. When you say your children come first, that becomes a rule — nothing gets built at their expense — not just an aspiration.
The interview is not what makes the constitution valuable. What makes it valuable is that the rules are specific enough to actually constrain behaviour, and honest enough to reflect what you would actually want the AI to do at three in the morning when you are not watching.
What usually goes wrong
People write an about-me file and then do not update it for six months. The AI session is oriented toward a version of their life that no longer exists. The biggest project from March is now done. The constraint that was urgent then has been resolved. The whole file is stale and the sessions are subtly wrong as a result.
The fix is treating the about-me file the way you would treat a monthly review: read it, update it, close it. Not every day. Enough that it stays true.
The other failure mode with the constitution is writing rules that are too vague to enforce. "Be careful with my data" is not a rule. "Never put my bank account number or my children's names into any chat session" is. The specificity is the point.
Why this is the most personal part of the system
The constitution is, in the end, a document about your values and your limits. What you are willing to trust a machine to do, and what you are not. What you are protecting, and from what.
That is why the document itself is the thing that has to be built by you, not handed to you. A template gives you a starting point. What goes in the template depends on what you are actually protecting — and that is a question only you can answer.
Drawn from chapter 5.4 of AI Magic 2033.
