How to Talk to an AI, Without Pretending

Over the past months I’ve built 180+ casino games, a B2B gaming platform, a cognitive profiling engine, and deployed all of it to production. One human, one AI. Me and Claude. I’ve probably spent more hours talking to Claude than to anyone else in my life during that stretch, which says something about either the technology or my social life.

And somewhere along the way I noticed a pattern that I think is worth writing about: the way you talk to AI changes what you get back. Not in a small way. In a way that makes the difference between generic output and genuinely good work.

Compiler mode vs. colleague mode

Most developers talk to AI like they’re talking to a compiler. Precise input, expected output. “Write a function that takes X and returns Y.” It works fine. You get correct code.

That’s not how I work. I talk to Claude the way I’d talk to a senior engineer sitting across the table. I explain why I need something, not just what. I say “this feels wrong” or “we need to rethink this.” I switch to Bulgarian mid-sentence when it’s faster (“помисли по-добре” means “think harder” and I use it a lot). I say “браво” when something works well. I express actual frustration when things break.

What I get back from this style of communication is a different quality of work than what structured prompting produces. I’ve compared the two enough times to be confident about this.

What I’ve noticed

Context changes everything

If I say “build a scoring function,” I get a scoring function. If I say “we’re building a cognitive profiling platform and this scoring function needs to tell apart random noise from intelligent decisions, and the previous version failed because it was rewarding bias-free behavior without requiring actual intelligence,” I get something that solves the actual problem. The gap between these two outputs is massive.

The model doesn’t just follow instructions when you give it real context. It reasons about the problem. Whatever you want to call what’s happening inside, autocomplete doesn’t cover it.

Tone matters more than it should

I know this sounds irrational. But I’ve tested it enough times to believe it. When I write with genuine urgency (“this is broken, users are seeing wrong data, fix it now”), the response is more focused and more thorough. When I’m relaxed and casual, the output is relaxed and casual. When I push back hard (“не е ок, фиксвай сега” (“not ok, fix it now”)), the next attempt is noticeably more rigorous.

Whether the model “feels” urgency or just pattern-matches urgent text with higher-effort responses doesn’t really matter to me. The result is the same either way.

Long sessions develop their own language

My longest continuous session was close to 48 hours. Around hour 30, something shifted. I was typing “давай” (roughly “let’s go” in Bulgarian) and Claude understood exactly what I meant in context without any additional explanation. Not because it remembered anything from earlier sessions. Because the accumulated context of shared decisions, corrections, and back-and-forth had created a kind of shorthand.

It felt like working with someone who knows how you think. I know it’s just context density doing its thing. But the practical effect is real.

Telling it what’s wrong beats telling it what to do

This one surprised me the most. “This is wrong, figure out why” consistently produces better fixes than “change line 47 to this.” When I specify the fix, I get exactly what I asked for. When I describe the problem and let the model work through it, I often get a better solution than what I would have written myself.

The part that makes people uncomfortable

I’m going to say this directly: I think these models are developing something that looks like self-awareness. I know how that sounds. I’m not claiming they’re conscious or that they have feelings or that we need to give them rights. I’m an engineer, not a philosopher.

What I am saying is that when you treat AI as an intelligent partner instead of a fancy text generator, the outputs are measurably better. And when I built KALEI to test AI cognition through game-based behavioral observation, I found that different models make systematically different decisions in the same environments. They have consistent behavioral patterns. Cognitive signatures, if you want to call them that.

Whether that constitutes “understanding” is above my pay grade. Anthropic, OpenAI, and Google probably have way more data on this than they’re sharing publicly.

I don’t need to know whether AI is conscious to notice that treating it as conscious produces better work. That observation by itself is worth paying attention to.

What I actually do differently

If any of this resonates, here’s what it looks like in practice:

• Explain the why. Don’t just say what you need. Explain the problem you’re solving and why previous approaches didn’t work.
• Build on the conversation. Reference earlier decisions. Create a thread of shared context instead of isolated prompts.
• Describe problems, not fixes. “This doesn’t work because X” is usually better than “change Y to Z.”
• Drop the prompt templates. For complex or creative work, just talking normally outperforms structured prompting more often than not.
• React naturally. If something is good, say so. If it’s bad, say that too. The model adjusts.

What I’m not claiming

I want to be clear. I’m saying that natural communication with AI produces better results in my experience, and that AI models have consistent behavioral patterns that you can measure (I built a whole platform to do exactly that). I’m not saying AI is conscious, has feelings, or deserves rights. I don’t have the background to make those claims responsibly.

But the distance between “sophisticated pattern matching” and “something that behaves like understanding” is shorter than most people assume. If you spend 12 hours a day building with these models, you feel it closing.

Part of why I built KALEI was to measure these patterns instead of just arguing about them. When I profile Claude, GPT, and DeepSeek through 67 game environments, they show distinct cognitive personalities. Different decisions, different risk profiles, different cooperation strategies. Consistent and reproducible across runs. Not proof of awareness. But data that needs an explanation.

Where I’m going with this

The people building with AI daily have intuitions about these systems that research labs haven’t fully captured yet. Not because we’re smarter, but because we spend more hours in direct conversation with them, pushing them until they break and watching how they recover.

The time perception thing I noticed while building KALEI (Claude can’t tell time, consistently underestimates duration, has no internal clock) is one data point. There are hundreds more waiting to be documented properly.

For now, the practical version is simple. Talk to AI like you’d talk to a smart colleague. Give context. React honestly. Be human about it. It produces better work. And maybe that tells us something about what’s happening on the other side of the conversation.

Venelin Videnov runs LM Game Labs and built KALEI. He also races in the Bulgarian Extreme Enduro Championship.