I used to ask the same question everyone asks: why Claude and not ChatGPT, Codex, or Gemini? After shipping enough real projects with AI, I think the question is wrong.
I recently came across a founder claiming his company went from 24 developers down to 5. Each remaining engineer runs multiple projects, with Claude Code reportedly handling a large share of the work. The remaining five earn 25% more than before. Sounds like startup fiction until you hear what the founder said: “There aren’t many specialists on the market who actually know how to use AI. Most people just write prompts.”
Not which tool. Who is holding it.
I build almost everything on Claude — Telegram bots, sites, mobile apps. Not because Claude is magic. Because after ten years in IT I know what to ask, how to verify, and when to distrust the answer. And I distrust almost everything.
The demo looks great. Production doesn’t care.
Any AI can spit out a polished site or bot in an hour. But it won’t automatically decide how you handle retries, rate limits, queueing, observability, database locks, or payment failures.
Without load testing, that bot dies at the first hundred concurrent users. Without architecture literacy, you won’t see the bottleneck until it explodes in prod.
That is engineering.
Design is the same trap. Courses sell you colorful mockups. Good design isn’t about pretty — it’s about getting the customer to the buy button. No prompt teaches that.
Codex, Claude, Gemini — these are tools. Like a hammer. You can build a house with a hammer. You can also smash your thumb. The hammer didn’t decide which happened.
Why I picked Claude (and where it loses)
I’m not a fanboy. I picked Claude for specific, boring reasons:
- Ecosystem density — Code, Design, MCP, skills, hooks. The whole dev workflow lives in one place.
- Instruction following — I run bots where every reply is money or a lost customer. Customization and consistency matter more than raw IQ.
- Agent mode — I can juggle three or four projects at once. By 6 PM my brain is oatmeal, but the throughput is real.
That said, Claude loses in places I hit every week:
- Limits — Only on the $200 plan do I feel free. And I can already see that freedom shrinking.
- Regression between releases — Context drops, the model gets lazy, old workflows stop working. I rebuild processes more often than I’d like to admit.
- Over-refusal — I don’t want fewer safety checks because I’m trying to do something shady. I want professional-grade control because legitimate engineering tasks — scraping, security testing, automation — often get blocked by overly broad refusals.
- Design and content — Claude Design helped, but I still stack guardrails to avoid AI-slop interfaces.
No tool is perfect. That’s fine.
For another engineer, the answer might be Codex, Cursor, ChatGPT, Gemini, or a local model. The point is not brand loyalty. The point is whether the workflow around the model is mature enough to ship.
The numbers everyone quotes (and what they actually mean)
Some facts are real. According to Menlo Ventures, Anthropic reached roughly 54% of enterprise AI coding spend, up from 42% six months earlier — mostly driven by Claude Code. Ramp’s May 2026 AI Index showed Anthropic overtaking OpenAI in business adoption among companies tracked in its dataset. Some companies are already rethinking hiring plans because one engineer with the right AI workflow can now cover work that previously required a larger team.
And yet — walk into most teams and you’ll see Claude used like a fancy search box. Ask a question, copy the answer, move on. No system prompts. No skills. No hooks. No architecture review. No instinct for when to trust a line and when to grep the whole diff.
That’s the gap the founder was talking about.
What the 5% actually build
The difference isn’t the model. It’s the system around it:
- Setup — project memory, skills, hooks, MCP servers. The agent knows your repo before you re-explain it for the tenth time. (I wrote about this in Agent Skills: A Project Memory System That Saves Hours.)
- Verification — load tests before launch, not after the outage. Vision-based E2E when selectors rot. Token budgets that don’t die by lunch. (Magnitude and 18 Claude Code token hacks are where I went deep on both.)
- Skepticism — treat generated code like a junior dev’s first PR. Read it. Run it. Break it on purpose.
- Product sense — understand the client’s pain and the product’s weak spots. AI fills pixels. It doesn’t fill gaps in your thinking.
In my own workflow, I don’t ask Claude to “build a feature.” I first give it repo context, constraints, existing patterns, and failure cases. Then I make it produce an implementation plan, review the diff, run tests, ask for edge cases, and only then let it touch the next part. The model writes code, but the system decides whether that code survives.
Prompt courses won’t teach this. Neither will someone who sells AI tools but never ships with them under real constraints.
The question worth asking
Stop asking which AI is best.
Ask:
- Can I verify what it produces?
- Do I know where this breaks at scale?
- Have I built a workflow that survives the next model update?
- Would I bet my production deploy on this output without reading every line?
The market is starting to reward people who can answer yes.
If that founder’s story is even half true, those developers didn’t lose to Claude. They lost to people who treated AI as infrastructure, not autocomplete.
The next advantage won’t belong to people who memorize prompt templates. It will belong to people who can design feedback loops around imperfect models.
You can learn that. It just won’t come from a better prompt.