As I’ve written before, the question of AI for writing code in the IT industry — I’ve settled that one for myself.
That programmers will resist is understandable. We’re all human and we don’t like it when something shows up and says “move over.” We don’t like being told we need to run faster and faster just to stay in place. These messages are uncomfortable, so our brain protects us — says let’s not waste glucose on these alarmists, we’ve seen this before, we’re smart and beautiful, everything will be fine.
But for me personally, the question is resolved. Over Christmas / New Year holidays, using Cursor + GPT-5.2, I managed to overhaul the architecture of a complex platform, add new functionality, cover it with tests, upgrade a bunch of packages, refactor, run security vulnerability tests and fixes, build an installer where one of the bash scripts alone is 2,000+ lines (it works and covers several important scenarios across both arm64 and amd64 architectures), add the ability to enable/disable modules, create a new monitoring module — and all of this across multiple GitHub repos, multiple tech stacks (Node.js / TypeScript, Erlang, React.js, Python, Bash, Docker) — and it works, and it’s in production.
That was my signal that the landscape has already shifted.
What to do next is more or less clear. And programmers aren’t going anywhere — quite the opposite. Especially those who are real engineers, or who can become engineer-operators.
When I wrote that I was doing vibe coding, people rightly commented that it wasn’t quite the right terminology for what I was doing. After all, I review at least superficially every operation, every change. I approve 80–90%. But the ability to quickly recognize what’s happening while staying in “flow” at AI speed, combined with knowing where those 10% are that you shouldn’t approve, where to say “what are you doing, that’s completely wrong, here’s what it should be” — that has to be there. That’s where a more unique and expensive-to-train biological neural network kicks in — 20+ years of experience in IT (technical, product, business), a PhD in Computer Science, experience building a product from code to $2M ARR in sales. OpenAI doesn’t manufacture that kind of “Murzik” yet.
What OpenAI / Anthropic produce in silico, without an experienced biological supervisor, either doesn’t build complex systems or builds them with issues.
(N.B. That said, “PhD in your pocket” has already been announced as a goal and is partially realized — so you should expect that by Q3 this year or next year, you’ll be able to summon as many PhDs, MBAs, Einsteins, and other Murzik Vasilyeviches as you want for specific tasks or ongoing work.)
But the industry has changed. Starting with GPT-5.2 and beyond, this is a tectonic shift where you either jump from plate to plate, or you wait for government or corporate evacuation later, with ready-made care packages.
You need to build processes differently, hire differently. Hire more “operators” who will plan architecture and implement AI-first from the start, only resorting to manual work when necessary. I suspect that in many cases, new AI-first teams should be staffed with juniors fresh out of university — just bring them in and say: this is how we work, this is how it’s always been. And those weird bearded guys who come in, perform ritual code reviews, and occasionally write code by hand in secret — those are mentor-dementors, don’t be afraid of them and don’t bully them, it’s just a religious ritual we have.
But personally, I’m not interested in building another Ciklum / EPAM / SoftServe, firing “old believers,” fighting groupthink, boiling the ocean, accelerating change, etc.
We already have a great product and team, and the fact that we’ll start shipping features 3–5x faster won’t dramatically change value for users, and definitely won’t change our commercial results short-term.
So I’m taking a small step back from using AI for code — giving the team time to catch up, digest the changes, and continue at a normal human pace.
Again, as I’ve written before, what’s more interesting to me — and what will truly take my breath away — is if AI can help me sell. If it can get me to millions in $ARR much faster for a relatively new product that’s almost unknown to the world.
Now that’s a challenge.
I actively started on this literally yesterday.
Since Cursor and I are already on “hey you, your mother!” terms, for sales and marketing tasks I decided to try Claude. For development, Cursor works perfectly for me, but I really like the ecosystem developing around Anthropic, I like the ASCII art in the terminal — basically, I’d been saving Claude Code for dessert, and now the time of Claude has come.
I started by:
- Googling and reading best practices on Reddit / Anthropic / Substack, and watching a few YouTube videos
- Looking into skills and several well-known repositories with marketing skills, but deciding not to plug them in yet
- Setting up a “BDSM” project and file system as a first approximation
- Connecting it to git
- Connecting data sources relevant to us (Google Analytics, HubSpot, Google Ads, MS Clarity, Ahrefs, Google Drive, Slack, JIRA) via MCP and APIs — almost nothing worked right away, but within 2–3 hours everything was up and running
- After connecting each data source and importing data, I asked Claude to generate a report with analytics and recommendations — this was the most enjoyable moment so far, receiving and reading those reports. There were a few unpleasant findings too, but their clear highlighting was genuinely great.
What can I say preliminarily, literally on day two?
It’s not exactly pure magic — it didn’t surface something I didn’t know about. Roughly everything in the reports was something I knew (or knew at some point but wasn’t holding it all in my head).
But there’s no fluff. No generic recommendations like “do lots of everything and lots will happen.” Everything is specific — about our traffic, our content, our product, our users, and our team. Clear, concrete facts and recommendations.
And here’s the real kicker — I knew all this as a founder/CEO who constantly swaps hats. My team doesn’t know all of it. Writing these reports would have taken me several days, then keeping them updated — same. Creating tasks and tracking their execution, making adjustments — that’s working weeks, months.
Nobody besides me has the full picture of where we are, where the market is, and where our opportunities lie in terms of sales / marketing / product. And accordingly, nobody had the ability to maximally effectively create tasks, review content, provide additional context on tasks, etc. — let alone creating initiatives or strategies. No, the team creates and does all of this, but either I’m the bottleneck significantly delaying tasks, or quality drops because the only person who could perfectly calibrate was me, having all the context and multi-disciplinary expertise.
But now that’s changing. I will no longer be the unique “Murzik”, the bottleneck founder guy. From what I can see, literally on day two of setup, this system is capable of significantly accelerating effective management and support of our sales and marketing work — and this is just the beginning. I haven’t even touched task execution by AI agents directly. I haven’t even fed the system my plans and proposals yet — I was curious to see what it would say based on raw data alone, and it aligned with my thinking about 90%.
To be continued — if anyone finds this interesting, I’ll keep writing about my experience building an AI-driven Business Development, Sales & Marketing department.