I Built an AI That Thinks Like Me. Here's the Gap I Had to Architect For.

Most AI assistants give you the average answer. I built one that gives you my answer. Here's the four-node architecture that makes the difference.

A colleague asked how I'd handle a delivery three months behind with no clear path forward. I fed the question to ChatGPT. It gave me a perfectly structured answer - five-step plan, risk mitigation, people involved communication. All correct.

All useless.

Because the generic answer wasn't wrong. It just wasn't how I'd actually handle it. It didn't account for the fact that I'd cut scope ruthlessly before adding people. It didn't know I prioritise shipping something over perfecting everything. It didn't capture that I'd rather have an uncomfortable conversation today than a crisis in two weeks.

After enough of these moments I realised: if I wanted an AI that could actually think the way I think, I needed to build and train it on how and why I think the way I do.

Here's what that took.

---

The Problem With Generic AI: Correct, But Contextless

ChatGPT is trained on the internet. It knows what most people would say. But "most people" don't have your constraints, your values, or your hard-won lessons.

When you ask it a strategic question, it gives you the average answer. The safe answer. The answer that sounds right in a vacuum but falls apart when it hits your actual situation.

That delivery question? ChatGPT suggested adding resources and extending timelines. Textbook project management.

I would've digested the intended outcome, ruthlessly cut features, shipped the non-negotiable outcome in four weeks with analytics, and dealt with the political fallout via phased deployments based on actual data. Because I've learned that scope creep kills more projects than tight timelines ever will.

That's not in ChatGPT's training data. That's in mine.

The gap between technically correct and actually helpful is where generic AI fails. And it's the gap I needed to close.

---

What "Thinking Like Me" Actually Means

It's not about mimicking my writing style. Voice is the easy part.

It's about capturing mode-of-thinking:

• When faced with scope creep, I ask: what's the non-negotiable outcome? What can move?
• When a process feels bloated, I strip steps until something breaks - then add back only what's essential.
• When someone asks for advice, I give them the uncomfortable truth with a path forward, not the cushioned version.

Those are principles. They generalise. They apply across contexts.

A lived story doesn't. "During a consulting project where scope exploded, we cut scope when..." is biography. It's specific to one moment. It doesn't teach the AI how to reason through the next situation.

So I built a system that captures principles, not just biography. Frameworks I use, mistakes I've learned to avoid, the questions I ask when I'm stuck.

That's what makes Ask-Marcus different to ChatGPT or Claude. It doesn't just know what I've written. It knows how I think. And it gets smarter every time I teach it something new.

---

The Architecture: Four Nodes, Built to Stay Grounded

Here's where most people get it wrong - they think the solution is feeding the AI more data. It's not.

It's about feeding it the right kind of data. And then separating every step so each one can do its job precisely, without bleeding into the next.

The system runs a four-node pipeline:

Node 1: The Router - Which version of me is relevant here? Am I answering as a strategist, a father, a fitness-focused person, or a tech lead? The router classifies intent and routes to the right reasoning frame. Without this, every answer comes back generic. The AI doesn't know which hat to wear.

Node 2: Decomposed Search - Queries five knowledge bases simultaneously: blog posts, raw learnings (principles and frameworks), voice profile, career experiences, reference materials. Returns chunks ranked by similarity to the question. Most systems query one source. This queries all five at once.

Node 3: Reranker - Rescores those chunks against the original question. Semantic search is great at finding related ideas, but it misses exact-term matches. Ask about "WSJF" and it might surface "prioritisation frameworks" generically, missing the specific learning that names WSJF directly. The reranker recovers that precision - and better retrieval means the reasoning node has accurate context to work from, not gaps it has to fill in.

Node 4: Reasoning - Takes the router's classification, the reranked chunks, and a reasoning frame. Produces a structured, analytical draft. Then voice synthesis converts that into how I actually talk.

Separating reasoning from voice means each step can be optimised independently. It's far better than my initial approach - one monolithic "write this in my voice" prompt that tried to do everything at once and did nothing well.

Each node solves a specific failure mode:

• Router: stops generic answers by activating the right version of me
• Decomposed search: stops missing relevant context by querying all knowledge bases
• Reranker: improves retrieval precision so the model has less room to fill gaps
• Reasoning separation: stops voice bleeding into analysis (think first, talk second)

That's the whole game.

And that delivery question? Ask-Marcus would cut scope and ship in four weeks - because it knows that's how I think, not because it guessed.

---

The Knowledge Base: Quality Over Quantity

The system stores knowledge in five namespaces:

blog-posts - Polished, self-contained thinking. Published content.

learnings - Raw capture of insights, frameworks, principles. This is the critical one. When I encounter a situation that reveals how I think, I capture it as a learning: title (what's the principle?) and content (explanation, examples, implications). This is what keeps the system from giving me that "all correct, all useless" feeling - it's not just retrieving facts, it's retrieving how I'd actually reason through the problem.

voice-profile - How I actually talk, encoded in markdown.

experiences - Role, org, period, key outputs. Career history for context.

documents - Reference material. Whitepapers, frameworks, SOPs, internal playbooks.

The learnings namespace is where the real magic happens - and the reason I said quality over quantity matters more than volume. A learning isn't a story. It's a principle that generalises. "Scope creep kills more projects than tight timelines" is a learning. "That time at Virgin when the project blew out" is biography. One teaches the AI how to reason. The other just gives it a fact.

The system is also proactive. It identifies gaps in my knowledge base before I notice them - thin categories, under-represented topics, questions that keep getting asked without a strong answer ready. Then it surfaces what to capture next, re-analyses coverage, and the cycle repeats.

The AI gets smarter every time I teach it a principle. That's the feedback loop, and it compounds.

---

Why This Approach Works

Most people think building a personal AI is about dumping more content in and hoping for the best. It's not.

It's about data quality first. Principles that generalise, not biography that doesn't. Frameworks, not just anecdotes. The right input shapes everything downstream.

Then it's about separation. Routing from retrieval. Retrieval from reasoning. Reasoning from voice. Each step solving one failure mode - not one monolithic prompt trying to do everything and doing nothing well.

The architecture scales without fundamental changes - more namespaces, more personas, more integrations are all additive. Every node is swappable: upgrade the router for deeper intent classification, swap the reranker for a custom model trained on your domain. And because reasoning and voice are separated, you can adjust how it sounds without touching the analytical core.

---

Where This Goes Next

The current version is live at marcushahnheuser.com/ask-marcus. The brain visual is at marcushahnheuser.com/brain.

What's coming:

• Multi-modal knowledge (files, images, transcripts, audio)
• API integration (real-time data, not just static knowledge)
• Agentic autonomy (proactive execution, not just responding)
• Feedback-driven prioritisation (track which questions get asked most, which answers rate highest)

Whether you're building your own version, want to embed this in a business, or just trying to understand how to architect AI thinking and not just AI output - this is the playbook.

Now open ChatGPT. Ask it a question only you can answer - something strategic, something that requires your specific judgment and experience. Notice where the answer is correct but useless.

That gap is yours to close. What's stopping you?