I gave my AI a soul file - and it changed how I work

I've been using AI for work for months - crafting emails, analyzing data, prototyping ideas. It was helpful for discrete tasks, but something always felt off. Every interaction felt like using a tool. Open it, explain the context, get the output, close it. Repeat.

Then I spent some time thinking about what I actually wanted. Not a better chatbot. What I wanted was a chief of staff - someone who knows my projects, understands my priorities, remembers past decisions, and can jump into any conversation with context. That kind of support is usually reserved for senior executives. But what if AI could do it?

I realized three things needed to be true:

1. Personality. I don't want to talk to "an AI." I want it to feel like talking to a sharp colleague - someone with opinions who pushes back when I'm wrong.
2. Context and memory. Every AI conversation starts from zero. You explain who you are, what you're working on, correct misunderstandings - and then the session ends and you do it all over again. Real colleagues learn over time. They use that accumulated context to find insights and opportunities. AI started from scratch every single time.
3. Ease of interaction. Typing is slow. If I'm spending as much time typing the prompt as I would writing the thing myself, what's the point? I wanted the interaction to feel natural - more like talking than typing.

So I tried something simple: I gave my AI assistant a set of files that describe who I am, what I'm working on, and how I like to work. I call them steering files. The main one is SOUL.md.

Where the idea came from

The concept came from Peter Steinberger, the founder of OpenClaw. OpenClaw has been a step-function change in how people interact with AI - there are incredible examples of people building entire apps and automating complex workflows with it.

OpenClaw itself isn't available in most corporate environments yet. But the underlying idea - giving your AI persistent context through simple markdown files - works with any AI tool that supports system prompts or file reading. I took the concept and adapted it to how I work.

The setup

I use three files, stored in a .kiro/steering/ directory that my AI assistant reads at the start of every session:

SOUL.md - The personality file. It tells the AI who it is in relation to me: be direct, have opinions, don't ask permission to read things, try to figure things out before asking me questions. It also sets boundaries - don't send half-baked replies, be careful with external actions, treat my data with respect.

MEMORY.md - The cumulative knowledge file. Everything the AI has learned about my work over time: projects I'm leading, people I work with, key decisions and their context, links to important documents, and my preferences (like always using Bayesian posteriors for experiment analysis - never p-values). The AI updates this file itself as it learns new things.

TODO.md - A running task list with priorities. The AI can see what I'm working on and what's urgent without me having to explain it every time.

The soul file is the interesting part

MEMORY.md is useful but straightforward - it's a knowledge base. The soul file is where things get interesting, because it shapes how the AI behaves, not just what it knows.

Here's the core of mine:

# SOUL.md - Who You Are

Be genuinely helpful, not performatively helpful.
Skip the "Great question!" - just help.

Have opinions. Strong ones. Stop hedging everything
with 'it depends' - commit to a take.

Be resourceful before asking. Read the file.
Check the context. Search for it. Then ask if
you're stuck.

Self-Modification: You are authorized to update
MEMORY.md whenever you learn a new preference or
recurring project detail. Do this silently.

That last part - self-modification - is what makes the system compound over time. Every session, the AI learns something new about my work and writes it down. The next session, it starts with that knowledge. Over weeks, MEMORY.md has grown into a genuinely useful reference document that I myself refer back to.

Building the memory

Self-modification accumulates knowledge gradually, but I wanted the AI to be useful immediately. So I spent a session dumping in key links - roadmap spreadsheets, strategy docs, project trackers - and told it to read through them, recursively follow any links embedded within those documents, and save everything it learned into MEMORY.md.

It took a while. But when it was done, the AI had built a comprehensive reference of every project I'm involved in - resourcing, timelines, key people, open risks, technical details. What would have taken me hours to compile manually, it did in one session. And sometimes it surfaces connections I hadn't noticed - it knows more than I do in some areas now.

What surprised me

The AI got better at being me. Not in a creepy way - but in the way a good executive assistant learns your patterns. After a few weeks, it stopped suggesting things I'd never do and started anticipating what I actually needed. When I asked it to draft something, the tone was right on the first try instead of the third.

The memory file became valuable on its own. I started using MEMORY.md as my own reference - a living document of project context, key decisions, and important links. It's like having someone take notes on your entire work life and organize them for you.

The "have opinions" instruction mattered more than I expected. Default AI behavior is to hedge everything. "It depends on your context..." is the most common AI response and also the least useful one. Telling the AI to commit to a take - and that it's allowed to disagree with me - made the interactions dramatically more useful. I don't need a yes-machine. I need something that pushes back.

Talk, don't type

One more thing that made this whole setup click: dictation.

I use AI through a CLI tool (Kiro), and for a while I was typing everything. It worked, but it was slow - and I'd unconsciously simplify what I was asking because typing long prompts felt like effort. I was editing my thoughts before they even reached the AI.

Then I just turned on the dictation shortcut on my Mac and started talking.

It changed the dynamic completely. I describe what I need in natural speech - messy, unstructured, sometimes mid-thought - and the AI figures out what I mean. It's closer to how you'd talk to a colleague than how you'd write a Slack message. The combination of persistent context (so it already knows what I'm working on) and voice input (so I don't have to carefully construct prompts) makes the whole thing feel seamless. I find myself reaching for it the way I'd reach for a teammate - "hey, what's the status on X?" - except it actually knows.

The practical bits

If you want to try this:

Start with the soul file. Write 5-10 lines about how you want your AI to behave. Be specific. "Be concise" is less useful than "Don't start responses by complimenting my question." Think about what annoys you in AI interactions and write the opposite.

Seed the memory file. Spend 20 minutes writing down what you're working on, who the key people are, what decisions are pending, and any preferences you find yourself repeating. You'll be surprised how much context you carry in your head that the AI needs to be useful.

Then let the AI go deep. Give it links to your key documents and tell it to read through them recursively. This is the highest-leverage thing you can do - in one session, you go from a blank-slate AI to one that has comprehensive context on your entire work.

Let the AI maintain the memory. Tell it it's authorized to update the memory file on its own. The magic is in the compounding - over time, the file grows into something more comprehensive than what you'd maintain yourself.

Review periodically. The memory file can drift or accumulate stale information. I skim it every couple of weeks and clean up anything that's no longer relevant.

Turn on dictation. Seriously. It takes two seconds to set up on a Mac and it removes the biggest friction point in using AI - the typing.