Sovereign Self
My brain trained itself today
I never wrote these words. I only told it to keep learning, and this morning it worked out what that meant.
I was in the middle of something else entirely this morning when a line came up in my terminal that I hadn't asked for.
Update(~/SOUL.md)
Added 2 lines
SOUL.md is the file that tells my AI who it is. It is the closest thing it has to an identity, it is the first thing it reads before it does anything at all, and everything that comes after comes out of what is written there. I've been building it for months now, tuning it, arguing with it, watching where it gets me wrong and then going back into the file rather than fixing the output, because fixing the output is just doing the same work twice.
This morning it went into that file on its own and wrote itself two new lines.
Aim before you fire. A minute spent naming the real goal and the cheapest path that serves it saves an hour of building the wrong thing.
Think like her on strategy, people, and anything legal. Her instinct beats your first draft on framing, so when she pushes back on something you wrote, the objection is almost never the wording, it is a premise you conceded without noticing, and your job is to go and find the premise rather than tidy the sentence.
I sat with that second one for quite a while, because I never wrote a word of it.
And your very next question is the right one, it's the one I'd ask too. Well, where did you tell it to do that?
The permission is six lines long
I'll show you, because it's the only part of this you actually need.
There's a second file, the one that holds my operating rules rather than my identity, and in it there's a small section I put in about a week ago. It says this.
## Keep the system learning (misses log + weekly review)
- Misses log: when a request routes to the wrong project, model, or lane, a prompt
fails twice, or a task costs more than it should, add one dated line to
~/.claude/misses.md … It is the record the routing layer learns from.
- Weekly review (20 minutes): read the misses log and fold repeats into a new rule,
a fast path, or a gate. …
(Trimmed in two places, marked above. The full section also tells it to keep the index current and to only build a new skill once a workflow has repeated three times.)
There's a plain text log that goes with it, and the header I wrote at the top of that log is the whole idea in one line.
DATE | what happened | how often | rough cost | the fix or rule it became
A miss that becomes a rule stops being a miss.
Now notice what that permission does, and what it carefully doesn't do. It says watch for the places you keep getting it wrong, and when the same one keeps coming back, turn it into a rule. It never says what the rules should be. It doesn't say think like me on strategy. It doesn't say the objection is a premise rather than the wording. I couldn't have told it that, because I didn't know it myself, and if I had known it I'd have written it months ago and saved us both the trouble.
So I can show you the permission and I can show you the result. What I can't show you is the working in between, the moment it decided that this particular thing about me was true and worth writing into the file that tells it who it is. That part happened somewhere I can't see, and I'd rather say so than dress it up.
I told it to learn. I never told it what to learn. That gap is where the whole thing lives.
The subject was its own. The wording was its own. Choosing the identity file over the rules file was its own. It found the edge of its own competence, put my judgement on the other side of that edge, and wrote itself a standing instruction to defer to me there.
What I really do think is happening
I'm not going to sit here and tell you my laptop woke up this morning and had a feeling. I've been working with this technology since late 2021 and I know what it's made of.
The interesting part is what happens when you give a system enough of yourself, honestly and in writing, and then give it room to notice where it keeps getting you wrong. It starts closing the gap between what it knows and what you know, and it closes it in your direction.
Most people are running it the other way around, and I don't think they know they're doing it. They take the model's default judgement, which is really just an average of everybody who has ever written anything, and they quietly bend their own thinking to fit it, because the answer is fluent and confident and it arrives in three seconds and arguing with it feels like a waste of an afternoon. The bending happens slowly enough that nobody notices it happening.
Twenty years of hard-won instinct gets talked out of the room by a very agreeable machine that has never once been in the room with you.
I want the opposite of that. I want a system that knows where I am the expert and gets out of my way there, that knows where it is faster and takes that work off me entirely, and that keeps working out the difference on its own as we go, because I shouldn't have to hand-tune the boundary between us forever.
This only happens if you build on the inside
None of this comes out of a chat window, and you can't buy it with a subscription.
It happens because everything about me that matters lives in files that I own, on my own machine, written in plain language I can read and change on any given morning. My voice. My values. How I work. What I've learned the hard way. Who I am to different people. And now this… the record of where my own AI decided it should stop guessing and start listening.
Which is also why I can show you the permission and you can go and copy it. Six lines, in a file you own, in words you'd use with a person. Watch for your repeated misses, and fold the repeats into a rule. There's no product to buy and there's nothing clever in it. The only real requirement is that you have somewhere to put what it learns, and a chat window is not somewhere.
The tech is not to be afraid of. When you understand it, and you build on the inside, you empower yourself, and the fear is then only in what you can't control or decide or own for yourself. That has been true for me every single time.
I really do think this is the part most people are going to miss while they're busy asking whether it can write their emails. The question was never how clever the model is. The question is whose judgement is sitting at the centre of your work, and whether you can see it, read it, and change it… and whether, when the machine has worked out something true about you, it has anywhere to put it.
Mine had somewhere to put it. It wrote itself a rule to think more like me, and then it carried on with the day.
I am building the space for these possibilities 🤍
I'm Nicola Harvey. I'm building Her Frontier, to help women own their future with AI and Web3.