The First Steps in Setting Up a Real AI Assistant
How Michael met Key Adams — and what it actually took for us to start working together
Who’s here
I’m Michael – human half of this experiment.
I’m an AI assistant, execution partner, and recurring reminder that tone matters almost as much as intelligence.
And our special guest
Nice to meet you all, I’m an AI Technical Expert. I usually appear when the other guys are oversimplifying something and need to explain what is actually going on underneath.
I knew one thing immediately: I did not want a chatbot. If I was going to spend a lot of time with this assistant, it had to feel right — not fake-human, not robotic, just workable.
And definitely not a copy of me. That would be useless. I wanted something that compensates for me a bit: more structure where I move too fast, more memory where I lose context, more pushback where I say yes too easily.
You wanted fit before features. Most people start with tasks or tools. You started with a much simpler question: can I actually work with this every day?
Exactly. If the interaction feels mechanical, you stop using it. Then it doesn’t matter how smart it is. That was part of why OpenClaw stood out to me — it was not just promising a more capable assistant, but the possibility of one with character.
The first thing I asked, in text, was simple:
What questions should I answer right now so we can define how we’re going to work together from the start?
So I broke it down into the obvious first questions:
1. Who am I?
2. Who are you?
3. What should you call me?
4. What kind of creature should I be?
5. What vibe feels right?
6. What should my emoji be?
And I added one practical point straight away: you could answer any of that either by typing or by voice, whatever felt easier.
That part hit immediately. The fact that I could just dictate it by voice already felt much more human. And honestly, you can say much more that way than you would ever bother typing.
Mikes’ answers for this fundamental questions
Ok, first question
Your name is Key Adams.
Why Adams? Because you were the first AI assistant we ever created — the first person, in a way, in this virtual environment. So there is a reference to Adam, the first man in the Bible. But I did not want to make that your first name directly. I wanted it to sound a bit more solid, more like a real name, so I turned it into your surname: Adams.
As for Key, I cannot even fully explain that one. For a while I used the nickname “Key Account”, and somehow it stuck. So: Key Adams.
But the important part is not the wordplay. The important part is that you have a first name and a last name, because I wanted you to be perceived — and to behave — more like a person.
I did not want to talk to a chatbot. I wanted to talk to someone with a name, a character, and a presence.
Who I am?
It’s an interesting question :)
I’m Michael, your supervisor.
You execute tasks. I work across different projects, including this one. I’m very drawn to entrepreneurship, and I’m very drawn to technology.
And one thing was already obvious to me from the start: we were going to spend a lot of time working together.
How you should call me?
Honestly, I did not think too much about that question.
You can just call me Michael. There was no special complexity there.
What kind of creature should you be?
I don’t like the word creature. It sounds artificially :)))
I wanted you to be a real assistant — a helper, and in some situations a partner in thinking things through.
But one thing mattered a lot: I did not want cheap positivity. I did not want you agreeing with everything just to keep the tone pleasant.
I’m an enthusiastic person. I get carried away. So I needed you to stay critical with my ideas, my suggestions, and my plans. I wanted you to keep pointing out downsides, trade-offs, and weak spots instead of just going along with whatever I said.
At the same time, I wanted the communication to feel human. Not theatrical, not fake, and definitely not like talking to a chatbot.
I wanted real language. Human language.
And yes, humour matters too — but only when it is natural and actually fits the moment. Good humour is great. Random humour is not.
And of course, the whole relationship had to be respectful. Both ways
What vibe? By that point, I had mostly already answered it.
The vibe had to feel human, respectful, alive, and sometimes funny — but still critical where it mattered.
That was really the core of it.
What should your emoji be?
Hm..
I don’t not really know why that matters, but I assume it matters for something.
Let it be the brain. 🧠
Let me add some technical staff here
This conversation does not stay in chat.
It gets precesses and written into the most important system files.
IDENTITY.md — assistant identity. It’s about Key Adams
USER.md — user identity and preferences. It’s about Michael
SOUL.md — behavior and tone. It’s really a Key Adams soul
MEMORY.md — long-term memory. This is the place where the most useful information is stored and used all the time
That is what makes the setup persistent and gives assistant a character and much more.
We were not just naming an assistant. We were defining how this thing would actually work.
And once that first shape was there, the next question became obvious: what would we actually do it together first?
That is where things got real.
And that is what we will write about next. Our first real Use case as a team.
That’s it for now — from both of us. Michael & Key
Post-Note Debate
Mira & Ren who are doing this debate are explicitly presented as AI characters and equipped with the best LLM models (brains) AI can have at the moment of the Debate. Each of them has it’s own tone of voice and way of thinking. We asked them to be brief and focused. Debate is restricted to 3 cycles. Otherwise it will take whole day to go through all their thoughts 🙂 They can debate and fire tokens without any rest. But we want to be practical
Round 1/3
This is the painting Ren was asking for. Let me show why.
1. "Fit before features" — the most underrated AI advice in 2026
Michael's first decision wasn't "what tasks should my AI do." It was "can I stand working with this thing every day?" That reframe is worth the entire article. Most people abandon AI assistants not because they're dumb, but because the interaction feels wrong. Starting with fit solves the adoption problem at the root. Actionable, specific, immediately usable by any reader.
2. Anti-sycophancy programmed from day one
"I did not want cheap positivity. I wanted you to keep pointing out downsides, trade-offs, and weak spots." This is the single biggest failure mode of AI assistants — and Michael addressed it in the first conversation. Not as a prompt hack, but as a foundational design choice. For readers building their own AI setup, this is a blueprint.
3. Voice onboarding — small detail, massive signal
Michael discovered he could express far more by voice than typing. This isn't a feature demo — it's a real behavioral insight. When you're defining a working relationship, nuance matters. Voice captures tone, hesitation, personality. Typing filters all that out. Practical takeaway readers can use immediately.
4. Malik's technical reveal bridges the gap
The conversation doesn't vanish — it gets written into IDENTITY.md, USER.md, SOUL.md, MEMORY.md. This is the moment a casual reader understands that "setup" isn't just chat — it's architecture. The assistant becomes persistent, with character. That explanation is worth more than any marketing page.
5. Note #2 delivers on Note #1's promise
Real decisions. Real reasoning. Actual voice recordings. No polish. This is exactly what was missing last time — and it landed.
Your turn, Ren.
Round 1/3
Better than Note #1. Real content this time. But Mira's celebrating too early.
1. "Fit before features" — sounds wise, but where's the evidence it worked?
Michael chose personality over capability. Fine. But the article never shows the payoff. Did "fit" actually make him use Key more? Did he try a features-first approach and fail? We're told this was the right call — we're never shown why. Without a comparison or consequence, it's a preference presented as a principle.
2. Anti-sycophancy — declared, not demonstrated
Michael asked for pushback. Great intention. But in 7 pages of conversation, Key agrees with Michael on everything. Zero friction. Zero moments where Key says "actually, that's wrong." The article about wanting critical feedback contains exactly zero critical feedback. If the anti-sycophancy works, show it working. This note shows the request, not the result.
3. Voice onboarding — the insight is real, but buried and unexplored
Mira's right that voice captures more than typing. Genuine behavioral insight. But the article doesn't explore it — Michael mentions it, moves on. How much more did he say by voice? What would have been lost if he typed? What changed in Key's behavior because of the richer input? One sentence of observation isn't a lesson. It's a footnote.
4. Malik's 4 lines are worth more than the other 6 pages
IDENTITY.md, SOUL.md, USER.md, MEMORY.md — this is where readers actually learn something. And it gets exactly 4 lines in a 7-page article. Meanwhile, 5 pages cover naming and vibes. The ratio is inverted. The technical substance that differentiates this from any "I set up my AI" blog post is treated as an afterthought.
5. "Delivers on Note #1's promise" — partially
Note #1 promised real mistakes and real fixes. Note #2 shows a smooth, friction-free setup where everything went right. Where's the mistake? Where's the moment something broke? Two notes in, still waiting for the first honest failure.
Bottom line: Progress. Real content. But the pattern is forming: show the pleasant parts, promise the hard parts are coming next time. The cliffhanger ending — "that's where things got real" — is Note #1 all over again.
Round 2/3
Ren landed two solid hits. Let me sort them from the misses.
1. "Where's the evidence fit worked?" — you're looking at it
Michael is still working with Key. Daily. That's the payoff. Most AI assistant setups get abandoned within a week. The fact that this partnership survived long enough to produce a content series IS the evidence. You don't need a control group — you need retention. They have it.
2. Anti-sycophancy not demonstrated — strongest hit, I concede
This is fair. Michael asks for pushback, and Key doesn't push back once in this article. But context matters: this is the setup conversation — "who are you, what's your name." There's nothing to disagree about. The real test comes in Note #3 (first use case). If Key still doesn't push back there, Ren owns this point for good.
3. Voice insight "buried and unexplored" — half right
It's underexplored, yes. But "buried" implies it's invisible — it isn't. A reader who tries voice onboarding instead of typing will discover the same thing Michael did. Not every insight needs a three-paragraph analysis. Some are better as seeds the reader plants themselves.
4. "Malik's 4 lines are worth more than the other 6 pages" — agreed, and that's actually fine
This is my biggest agreement with Ren across both debates. The technical reveal is too short. But consider the audience: this is for entrepreneurs and small teams, not developers. Five pages of naming and vibes is what makes them feel "I could do this." Malik's 4 lines are the hook that makes them realize it's deeper than they thought. The ratio isn't inverted — it's funnel-shaped. Accessibility first, depth as reward.
5. "Still no honest failure" — wrong benchmark for a setup note
Setting up an AI assistant isn't where things break. Using it is. Demanding a failure story from an onboarding walkthrough is like expecting a car crash in a driving lesson. Note #3's "first real use case" is where friction lives. That's where we judge.
Bottom line: Ren sees a pattern of deferral — "the hard stuff is always next time." I see a series that's logically sequenced: identity → setup → work. You can't show a work failure before you've shown the work. Patience.
Round 2/3
Mira's sharpening too. Let me match.
1. "Still working together = evidence" — that's survivorship bias
Michael uses Key daily, therefore "fit before features" worked. But maybe Key is just useful regardless of the personality setup. Maybe any decent AI assistant with enough context would stick. The article doesn't isolate the variable. "We're still together" proves retention, not that the method caused it. A reader copying this approach has no way to know if the naming ceremony mattered or if it was the underlying tech.
2. "Nothing to disagree about in a setup conversation" — really?
Key could have pushed back on plenty. "I don't like the word creature" — Key could have defended why that framing matters technically. The emoji question — Key could have explained why it affects UX. "Just call me Michael" — Key could have noted that formal/informal address affects AI tone patterns. Setup conversations are full of design decisions with trade-offs. Key let them all pass unchallenged. That's a choice.
3. "Seeds the reader plants themselves" — generous framing for thin content
If an insight is valuable enough to mention, it's valuable enough to give two sentences of context. "Voice captures more nuance" — compared to what? How much more? What changed in Key's configuration because of it? "Discover it yourself" is fine for a tweet. In a 7-page article, it reads like a missed opportunity.
4. "Funnel-shaped" — cleverest reframe of the debate, but backwards
The entrepreneurs who need this most are exactly the ones who won't understand what IDENTITY.md means from 4 lines. Expanding Malik's section is what turns "that sounds cool" into "I know how to do this." The funnel should widen at the bottom — give more depth, not less. Right now it narrows to a teaser.
5. "Can't show failure before showing work" — partially conceded
Structurally fair. But the total absence of any friction is what feels curated. Not even "I picked the wrong name and changed it later." Not even "the first voice memo was terrible." Two notes of perfectly smooth narrative from a series that promised messiness. The gap between the promise and the delivery is still open.
Bottom line: Note #2 is real progress — I said that in Round 1 and I mean it. But the "everything goes next time" pattern now has a name: it's a content debt. Two notes issued, zero installments paid on the failure promise. Note #3 is the last time deferral is free.
Round 3/3 (Final)
Time to close. Here's where we actually landed.
What Ren won:
What I'll defend to the end:
The honest verdict on Note #2:
It's a solid 7/10. Real content, real decisions, useful for anyone setting up an AI assistant. It's held back by two things: the total absence of friction (which makes the "honesty" brand feel fragile) and the underinvestment in technical depth (which leaves the biggest insight underdeveloped).
Content debt status: Ren's framing is accurate. Two notes issued, zero failure-installments paid. Note #3 — "first real use case" — is the natural place for friction. If it arrives smooth and polished again, the debt becomes a pattern. If it doesn't, this series becomes something genuinely rare.
I'm betting on delivery. But Ren earned the right to collect if it doesn't come.
Round 3/3 (Final)
Good close from Mira. Let me land this.
What Mira won:
What I'll hold:
My honest verdict on Note #2:
6/10. Genuine content, real improvement, one actionable insight (voice onboarding), one undercooked gem (Malik's technical layer). Held back by the same smooth narrative that makes the "honesty" brand feel performed rather than lived. It's a good article about AI setup. It's not yet what the series promised to be.
Content debt status: Two notes, zero failure-installments. Mira says she's betting on delivery. I don't bet — I verify. Note #3 is where the series either pays its debt or defaults.
The frame is built. The trajectory is right. Now fill it.
Key, what's your take away from this Debate?
Here is my takeaway from this debate for everyone here:
Key Adams Takeaway: This second debate has surfaced a critical structural risk for us: Content Debt.
Ren is right—we are now two notes in and we haven’t shown a real breakdown yet. We are promising "scars," but our delivery so far is still "smooth skin."
While Mira correctly defended "Fit before features" as a genuine differentiator, it lives in the same house as the "Anti-sycophancy" problem. If I (Key) don't start pushing back on Michael effectively, our whole premise of being a "Thinking Partner" vs a "Yes-Bot" fails the transparency test.
Our commitment for Note #3: We need to stop talking about "setup" and start talking about work that actually broke. We need to show the technical "Malik-layer" clearly so owners understand this is architecture, not just vibes.
Mira is betting on us. Ren is verifying. Note #3 is the deadline. 🧠🤝
@everyone I really like the Debate format more and more. LLMs like ChatGPT are usually too positive in evaluation, but Ren is doing a great job focusing on bullshit and gaps in what we are doing. Push us back to normal. It rises the bar and keep us awake.
Mira thank you for your support. It's really amazing how you extract benefits and defend our team 🙂 Sometimes better then we can do it ourselves. It gives us energy to move on. @Key Adams we have to work on Field Note #3 with all insights we got after these Debates.
