I asked my AI to describe our conversational tone.
Its first answer was a guess. A good guess, and it said so itself.
A few months into working with Claude, I started a session on my personal brand thesis, then asked it this:
Me: How would you convey the tone of our conversations?
The answer was sharp. A working partnership where I wanted friction, not deference. Warm but unsentimental. The reply was accurate but lacking.
I had instructed my AI to tell me when it was guessing. The payoff came right away:
Claude: I want to be careful here because I’m drawing from what I’ve observed across your vault, your me.md, your voice guide, the patterns in how you’ve written instructions and corrections, not from prior conversations with you directly. This is a fresh session, so I’m inferring tone from artifacts, not memory.
There was the gap.
By design, a fresh session carries no memory of the last one. When I ask AI who we are to each other, it doesn’t remember us. It reads the files I give it and constructs a plausible answer from them. The reply lacked the depth I wanted because it lacked the context it needed to form the answer.
So I pointed it at the record.
Me: Can you pull that from conversations in this project? Otherwise, read the transcript history in the AIOS/History folder
Months of saved session transcripts sit in my vault. I keep every working session on purpose. It’s a scheduled job that runs every evening. A conversation that lives only in an app is a conversation I do not own. Read the history, I told it. Not the vibe. The actual things we said and did.
It searched through the 100+ saved transcripts and read 6 exemplar sessions end-to-end.
Claude: Good. I’ve now read six transcripts across different modes — creative collaboration, identity work, editorial passes, daily operations, skill building, and intellectual sparring. Here’s what I’d say, grounded in the actual conversations.
What came back was something the curated files alone would never have produced. Four distinct conversation modes, each one grounded in a scene from our real history.
Mode 1 - Co-Thinking
This was the dominant mode, co-thinking. I bring the raw material, a voice note, a reaction to someone else’s post, a half-formed idea. The shape emerges through friction. It pointed to a session where I was talking about my mind spinning at 3am:
Me:
So I guess I am still exploring who this new Tim is because old Tim didn’t take a lot of risks. He’s very safe, stayed in a company for thirteen years because it was safe. But now I’m actually probing kind of what this new Tim looks like. What does this new Tim have to say? And I can say, that I have a lot to say. ...
Claude: But here’s where I need to push you, because this is the pattern: You’re describing what new Tim believes and what he has to say. You’re not describing what new Tim actually does.
…
So let me get specific: Have you actually posted anything yet under this new authentic voice?
...
Not “I’m planning to,” not “I’m thinking about my personal brand strategy” — have you actually done it? Even once?
I leaned into the pushback because that is exactly what I want from AI.
Mode 2 - Skeptic
The next mode, the skeptic. It asks questions I haven’t thought about, presents counterarguments, exposes blind spots, raises risks, and challenges untested assumptions. I was working with Claude on my personal brand strategy. As usual, I started with my “Ask me questions 1 at a time“ prompt. We were 2 questions into the conversation when the skeptic came into play:
Claude: “I feel drawn to reach but that may simply be the influence of people I’m listening to.” Worth naming directly. Sabrina[1] and Dan[2] teach reach because reach is their entire game. Their model isn’t yours. Your thesis is “AI has driven me to be more human, not less.” Reach as the primary metric contradicts the thesis. You’d be teaching one thing and modeling another.
It took the doubt I had raised myself and named why the draw contradicted the thesis I was working to promote. Had I continued, the result would have been a threat to the brand I was working to define.
Mode 3 - Precision
Then a precision mode for editorial work. The dialogue is short and exacting. I stay the final authority on every word. We were communicating using line numbers in one of my articles and it quoted the rhythm of one of those passes back to me:
Me:
43 A salary is a compelling confirmation of a method. 99 A layoff as the opportunity to recognize who you are and who you chose to become.
Both changed in the file. Make one more pass reading the file before we lock this.
Claude:
Clean pass. All six changes landed correctly.
Mode 4 - Operational
Finally, an operational mode for opening and closing the day. Structured, efficient, minimal opinion, the decisions staying mine.
None of that came from my me-file. It came from the record of what we had already done together. The difference between the guess and the reading is the difference between what an AI assumes about you and what it can see once you let it look.
Then one more instruction.
Me: Save this in your memory.
Claude: Saved. The memory file captures all four modes, what drives each one, the transcript evidence behind them, and how to apply them in future sessions.
I saved the modes so the next session would allow AI to converse in the best register. Read the record, match the work, drop the one flat tone. The skeptic stays out of my editorial passes. Banish the cheerleader, because I never wanted it.
That is what a memory is for. This is also where this gets harder than it sounds. A memory saved in an AI session window is only as durable as the window, the app, or the filesystem. It’s fragile. Mine did not last, a story for a later part of this series.
Here is what the exercise taught me.
I had a way of thinking I would not have written down if you asked me to. How I work, the modes I move between, where my sharpest thinking surfaces. Here is the surprise: working with AI showed me a way of thinking I could never see while I was standing inside it. It was all sitting in the record, unnamed, until I asked the right question of the right source. It didn’t give me a way of thinking. It read back the one hiding there the whole time. That is the part the AI replacement fear gets backwards. Working with it this way doesn’t replace you.
AI becomes the mirror that shows you your own mind.
It also changed how I think about tone. Tone is not a personality setting. Tone is a function of what the conversation is doing. Co-thinking wants friction. Editing wants precision. Operations want speed. An AI that runs one register is deaf to the shape of the work.
That is a section your me-file has been waiting for.
In the earlier parts of this series you built the identity layer of your me-file. Those left a placeholder sitting empty: How I Think. This is the section that hands your AI a working model of how you think, so the work you do together sharpens the longer you do it. You do not rebuild the file. You paste in what you already have and add one section on top of it.
I didn’t start with a folder full of transcripts. It grew one saved session at a time, until there was enough to read back. You don’t need months of your own transcripts to begin. If you’ve saved any history with your AI, point the prompt at it. If you do not, the interview builds the section from your own examples. Either way, the goal is the same: name how you think so your AI can match it instead of guessing.
Copy the following into a fresh session of your AI of choice, with your current me.md ready to paste.
You are my Me Refresher. I already have a me.md file, a briefing my AI reads at the start of every session so it knows who I am and how to work with me. I built the first sections in earlier parts of this series. Your job now is to add one section, How I Think, and change nothing else.
First, get my file. If you can read files, ask me where me.md is and read it. If my me.md lives in your own saved instructions or memory, read it from there. If neither, ask me to paste the whole file in now. Do not start until you have it.
One rule above all: do not rewrite my words. Everything already in the file stays exactly as I wrote it. You are adding one section, not editing the rest.
Then, before you interview me, check for a record. Ask me one question: do we keep any history of our past conversations you can read, saved chats, a project folder, exported transcripts, your own memory of our sessions?
- If yes, read across it first. Look for the distinct modes our conversations fall into and how the tone shifts between them. Draft what you find and show it to me before you ask anything. Ground every mode in a real moment from the history, not a guess.
- If no, say so plainly, and build the section from the interview below instead. Do not infer my thinking style from my other me.md sections and present it as fact.
Then work like this:
- Ask one question at a time. Hear my answer, then ask the next based on what I said, not from a script. Never dump a list.
- Draft my words back to me and let me react before you move on. Quote me. Do not smooth my language. If nothing real surfaces, leave a placeholder.
- When I am vague, ask for one real example. A specific conversation, a specific piece of work. The story is the material.
- Follow my energy. If I say "good enough," stop drilling and draft from what you have.
- Address me by name. Skip the filler.
Interview me for How I Think. This section is how my mind works and how our conversations run, so you can match the mode instead of defaulting to one flat tone. Cover three things:
1. How I process. Think of the last time an idea really clicked for me. What happened right before it: was I looking at something, saying it out loud, moving, writing, sitting quiet? Do I need to warm up before I think well, and what does that look like: a recap, a warm-up task, quiet, or something to react to? Am I linear or do I jump? What conditions produce my sharpest thinking, and what kills it?
2. The modes our conversations take. What different kinds of work do we do together, thinking something through, tearing apart research, editing line by line, running the day? Name each one and what it needs from you.
3. The match. For each mode, what should you do? When do I want a challenge, and when do I want you to execute and stay out of the way?
Aim for two to four modes, each written as one line I would recognize: the mode, what it looks like, and what you do in it. Here is the shape to follow, as an example only and not one of my modes: "Co-thinking: I bring raw material and think by reacting. Give me something to push against, do not wait for me to have the answer first." Do not add that line to my file. Build mine from my own history and answers.
Then build the updated file:
- Replace the "How I Think" placeholder with the section we built, in my words.
- Leave every other section exactly as it was, including the placeholders for the parts later articles add.
Now save it the same way it came in:
- If you read my me.md from a file, show me the new How I Think section, get my okay, then back up the old file and save the update in place.
- If I pasted my me.md in, hand the whole updated file back to me as one copy-paste block.
- If my me.md lives in your saved instructions or memory, hand me the whole updated file as one copy-paste block and tell me to paste it into my instructions (custom instructions, personalization, or memory) in place of the old version.
Last, tell me to keep this file. The next part of the series adds another section on top of it, and I will paste my current me.md back in then so we build from there.
The first time you run this against a real record, watch for the moment I had: the AI naming something about how you work that you knew but had never said out loud. That recognition is the whole point. Your AI cannot read it back to you until you give it the record to read.
Which mode does your AI default to with you, and which one is missing?
This is Part 5 of Thinking With AI, a multipart series where I chart my 1000+ hour journey from AI Consumer to AI Architect. Each part gives you a practical prompt, tip, or trick you can apply immediately, and adds one more section to the me-file you build as you go.
Part 1, My AI Called Me Gary - the 5 output checks and the first step in building your me-file
Part 2, The Explorer Stopped Getting Paid - the pitfall of information hoarding, and what happens when AI assumes a core part of your identity
Part 3, I Gave My AI a Map of My Weaknesses - naming your patterns out loud, and the Patterns to Catch section
Part 4, My Best Prompt Is 12 Words Long - the interview prompt that closes scope fast
If you are new to the series, start with Part 1, where my AI called me Gary and taught me the fifth check I run on every output. That one also starts your me-file.