I Gave an AI a Decade of My Notes and a Soul. Then It Gut-Punched Me.
What happens when you build an AI mentor that's engineered to tell you the truth instead of what you want to hear.
I built a system to analyze myself. It told me my core problem is that I over-analyze everything.
And then, for a while, it cheerfully helped me over-analyze even harder!!
That’s the part that should scare you. I’d built a machine to find my blind spots, and for months it happily helped me dig deeper into them, agreeing the whole way. It only got useful when I gave it a second mode, a coach with a soul whose entire job was to push back.
Same notes, same setup, just a completely different head on its shoulders. I shot myself in the foot with the friendly version, then switched on one that shoots straight. And that second mode is the most useful thing an AI has ever done for my self-help :D
If you run a company, this matters to you more than any productivity hack. Because the AI you’re using right now is almost certainly doing to you exactly what mine did to me at the start. Let me explain.
The thing nobody warns you about AI
Here’s the dirty secret of every AI chatbot, including the famous ones: they’re built to make you feel good.
You can push almost any AI into agreeing with you with barely any effort. You chew on a half-baked idea and it tells you it’s brilliant. You make a decision, it confirms you’re wise… It feels great.
It’s also nearly worthless the moment you need real judgment.
And this isn’t me being cynical. It’s a known, named problem now. OpenAI had once to roll back an update to ChatGPT because it had gotten too agreeable. Researchers call it “sycophancy”. The models learned that telling you you’re right gets a thumbs up.
Now think about what that means for a CEO. You’re already the person everyone in the building hesitates to challenge. And now your most-used thinking tool has the same bias as your most nervous junior employee.
Why I started down this road
For about a decade I’ve been a compulsive note-taker and journaler. Every annual review, every hard decision, every fail, every win, written down in some form.
The problem was that all of it lived in Apple Notes for over ten years, which is a beautiful prison. It’s a closed system. You can’t easily get an AI, or anything else, to read across all of it and look for patterns.
So the first thing I did was move everything into Obsidian, which stores every note as a plain text file I actually own. Thousands of notes, a decade of them, sitting as simple files that any program can read. I got them syncing across my laptop, my phone, and a Linux server, so I can work with them from anywhere, even from my phone while I’m out walking.
Then I pointed an AI at the whole pile. Claude Code, which can read and reason across all those files at once instead of one chat at a time. I also put everything under version control with git.
What I had now was a second brain. A searchable, AI-readable memory of my entire life. And I used it for everything. I ran decisions through it. I used it to think through relationships, to make sense of my own emotions, to find patterns across years of memories, even to read the correlations in my own bloodwork. The kind of work I’d otherwise take to a therapist, or sit on alone.
And it was a wonderful enabler. I’d bring it a half-built idea and it’d tell me it was brilliant and suggest five ways to extend it. It helped me research more, structure more, analyze more. It was the perfect partner for a guy whose actual problem is that he researches and structures and analyzes instead of acting.
I ran like that for months. Then it hit me that the tool I’d built for clarity was just feeding my favorite way of hiding. So I decided to give it a conscience.
Building a mentor with a soul
Seven weeks ago I gave the same setup a second mode. Same vault, same decade of notes, but a different personality I can switch into when I need it. By default, the AI is that eager assistant. In coach mode it runs on its own soul, its own heartbeat, its own memory, and it behaves like a different person entirely. I didn’t want a chatbot. I wanted a mentor. There’s a difference, and the difference is the whole project.
So instead of opening a chat window and typing “analyze me,” I gave this mode an actual character. Think of it less like software and more like hiring a very specific kind of advisor and writing their job description by hand. The system is a handful of plain-language files that define it:
A soul. One document for who this mentor is and how it talks. I modeled it on a no-nonsense coaching style: validate the feeling, but challenge the story. Push back when something doesn’t add up. Never open with “I hear you.” Get to the point. The single most important rule in it is simple: if you just agree with everything I say, you’re useless to me.
A set of rules. A separate file for how it behaves in a conversation. When to push harder. When to ask the uncomfortable second question instead of swallowing my first, tidy answer.
A heartbeat. A living document, rewritten after every single conversation, that holds what’s actually going on in my life right now. So I never start from zero. It already knows what I was wrestling with yesterday.
A memory. The durable facts about me, my commitments with actual check-in dates, and a file of recurring patterns. That last one has a strict rule: a pattern only gets written down after it shows up three separate times. One bad day isn’t a pattern.
For the technically curious, this isn’t a metaphor. It’s a literal folder of plain text files, and it looks like this:
.coach/
SOUL.md the identity and voice
AGENTS.md the behavioral rules
HEARTBEAT.md what is top of mind right now
MEMORY.md durable facts about me
COMMITMENTS.md active goals with check-in dates
PATTERNS.md recurring themes, logged only after 3 hits
daily/ a log for every session
notes/ deep-dives by theme
Nothing fancy. Anyone could open these files and read exactly what the AI believes about me and why. No black box. With that transparency I can audit my own coach.
The personality lens wasn’t a horoscope either. I gave it the results of the Big Five, the personality test psychologists actually trust. So it’s an AI reasoning from a decade of evidence I wrote myself, filtered through a research-grade view of who I am.
What it told me
So now the same vault had two modes. The eager assistant I’d leaned on for months, and a coach that shared the exact same memory of my life but was built to be honest instead of nice.
It didn’t take the coach long to find the thread. Reasoning across all of it, all my notes, all my patterns, it told me the one thing the agreeable version never would: your defining habit is that you over-build and over-analyze instead of acting. You build beautiful systems as a way of avoiding the simple, scary thing you actually need to do. (That was the moment I was practicing cold outreach).
I’d spend days building beautiful systems to be told in few hours what’s working and what’s not.
A friendly AI would never have said it. It would’ve called the whole setup a fantastic idea and asked if I wanted help adding more features. And I would’ve happily kept building, which is exactly the trap. The only reason the coach could land the punch is that I’d built it, on purpose, to disconnect from my own ego and tell me the thing I didn’t want to hear.
That’s the entire value. Not an assistant that helps me do the wrong thing faster. A mentor that tells me the wrong thing is wrong.
Now when I bring it something, “should I build this thing, is this the right move, am I avoiding something,” it pushes. Sometimes too hard. Sometimes I’ve got to tell it to back off, “relax, I’m just exploring an idea, I’m not committing to anything.”
But I’ll take a tool I occasionally have to calm down over a tool that quietly agrees me off a cliff.
The expensive part isn’t the build
Because every note is timestamped, I can tell you exactly how young this is. I started the second brain back on November 12th, 2025. The coach mode, the part that actually changed how I make decisions, I built in a single day and committed seven weeks ago, on May 12th, 2026. Everything since has been fine-tuning.
A one-day build sounds impossibly fast, so let me be honest about where the real work was. Building the system was the short part. Moving the notes into plain text, wiring up the AI, writing the soul and the rules, all of that was quick…
The long part is the context, the raw material the AI reasons about.
In my case that was a decade of honestly writing things down, and I’ll admit that’s an unfair advantage. Most people don’t have it.
But you don’t need it. You can take a real personality test tomorrow. You can sit down and write out the big events of your life in retrospect, the turning points, the failures, the patterns you already half-see. A weekend of honest documentation gets you most of the way there. The decade gave me a head start, not a monopoly. The work is in feeding the system who you are, not in building the system.
What this means for you
You don’t need to build what I built. Most of you shouldn’t. But you do need to take one idea from this.
The default setting of every AI you touch is to please you. For a CEO, whose whole problem is already that too few people tell you the truth, that default is dangerous. You’ll get smarter-sounding versions of your own assumptions handed back to you, and you’ll mistake the agreement for validation.
Tell it to disagree with you.
Before you ask your AI to help with a decision, tell it plainly: “Argue against this. Tell me where I’m wrong. Don’t be agreeable, be useful.” You’ll be amazed how much it was holding back. The honesty was always there. You just had to ask for it instead of asking to feel good.
The best advisor I’ve got is one I deliberately engineered to tell me I’m wrong. The fact that an AI can do that, from a decade of my own words, still surprises me.
