Intent is the new craft: 29 days of data, 10 years of AI work

Build is cheap. So what is the craft?

If a weekend with the right AI tools can ship a working product, the rest of us — twenty-year veterans, weekend hackers, juniors who learned with Claude — are all standing in front of the same question. What kind of work still belongs to humans? Where does seniority earn its name? And how do we know we did good work when the AI did most of the typing?

I have been building software through three AI eras: deep learning before LLMs existed, GitHub Copilot when it autocompleted code, and now Claude Code as a teammate I run multiple of in parallel. Last month I logged 29 days of dense usage — 3,712 messages, 275 commits, 238,530 lines added. The numbers are not the point. The point is that the work feels different now, and I want to think out loud about why.

Three eras, one trajectory

In 2015, AI in production meant deep learning — your own model, your own labeled data, your own GPU cluster. The work was 80 percent data engineering and 20 percent training tricks. If you wanted AI in your product, you built it.

In 2022, Copilot landed and turned coding into autocomplete on steroids. ChatGPT, GPT-4, and Gemini followed. Each wave pulled a little more cognition out of our heads and into a chat box. The mental model was helper, not partner.

In early 2025, Claude Code shipped — the first tool that did not feel like an assistant. It read files, ran commands, wrote tests, and pushed back. The mental model flipped from helper to crew member.

The trajectory is one direction: from "you train the model" to "you steer the agent." The current wave is the strangest one yet, because the agent is not just typing for you. It is also reasoning, deciding, and sometimes making the wrong call.

What intensive use actually looks like

Before the takeaways, here is the shape of one heavy month. I include it so the rest of this essay does not feel like speculation.

Two patterns stand out. The first is constant delegation — sub-agents handle about a third of the workload, and 35 percent of sessions overlap in time, meaning multiple Claude windows running at once. The second is hard pushback. The report flagged 38 dissatisfied turns against 289 satisfied ones, and one pushback session uncovered 23 real defects in a release that initially looked clean.

This is what working with an AI crew member looks like at full intensity. It is not chat. It is operations.

The first surprise: juniors might be better at this

Twenty years of full-stack work taught me how things should be built. That knowledge is now a tax.

When Claude Code suggests an architecture I would not have picked, I argue with it. Half the time those opinions slow me down — the AI's suggestion was fine, and I burned tokens defending mine.

A junior does not have those opinions yet. A junior reads the suggestion, runs it, sees if it works, and moves on. The result lands faster.

Senior craft still matters most where the AI is wrong — non-obvious failure modes, security boundaries, one-way doors. But the daily 80 percent of code is no longer a senior's domain. It is a junior's domain plus a careful reviewer.

The trap for seniors is ego: defending your way of doing things when the AI's way is good enough. The trap for juniors is verification: shipping what the AI gives you without proving it works. Whoever fixes their trap first wins.

The CTO job moves from headcount to intent translation

For most of the past two decades, the CTO job was three things: manage headcount, manage tech debt, and ship on time at acceptable quality. Those tasks survive, but they are no longer the daily work.

The new bottleneck is intent precision. When the AI fleet has clear intent, a six-person team can ship in five weeks. When intent slips, the same team wanders for a quarter. I lived this on a small AI-driven logistics startup that closed a Series B on the back of an unusually fast first product. The velocity was not magic. It was the result of relentlessly translating fuzzy human intent — what the customer actually wants, what the business actually needs — into specifications a fleet of AI agents could run with.

The other half of the new job is verification with evidence, not vibes. The highest-impact habit I have found is a single Korean phrase, "추측없이 꼼꼼하게" — "thoroughly, without guessing." It is a pushback marker. Every time I use it, the AI re-runs tests, checks logs, and returns evidence instead of claims. The English version looks like this:

verify this without guessing — re-run the tests,
inspect the logs, and return evidence (file:line),
not claims.

The new CTO move: refuse to accept "done" until the system proves it.

Vibe coding is will plus intent, nothing else

I do not understand every line in those 238,530. I do not need to. What I needed was three things: a clear picture of what should exist, the will to keep going when the AI was wrong, and a verification habit that caught lies before they shipped.

That is the whole job.

Most people who try vibe coding stall on one of the three. They have unclear intent and ship a product nobody wants. They lack will and abandon at the first compile error. They skip verification and watch their prototype rot in production.

The good news: will and intent are not gated by a CS degree. A logistics expert with precise intent and the will to push back on bad output can ship a logistics product. A founder with a sales background and clear intent can ship a sales tool. The technical knowledge is no longer the bottleneck. The clarity is.

Solo founders will rise. Most will not make money.

Build is cheap now. That is the easy half.

The hard half is that anyone can ship, so shipping is no longer differentiation. The moat shifts to two things: domain depth and distribution.

Domain depth means you understand a market well enough to know which product is worth building. Anyone can spin up a SaaS in a weekend. Knowing which SaaS the market actually pays for takes years of being inside that market. That is a moat AI cannot give you.

Distribution means you can put the product in front of people who will pay. This is the part LLMs cannot do for you. They cannot run cold outbound, attend industry events, or build a five-year reputation. They can write the email, but not earn the trust.

The new wrinkle is AEO and GEO — Answer Engine Optimization and Generative Engine Optimization. As more buying decisions start with "which X is best," asked directly to an AI, getting the AI to recommend you matters more than ranking on Google. I wrote about this in How I made the AI recommend my app.

Without domain depth, you build the wrong thing. Without distribution, nobody finds it. Without AEO, the AI recommends a competitor. Solo founders rise; most do not collect.

The end of pyramids

The 20th-century company was a pyramid. Mass production at the bottom, middle managers in the middle, executives at the top, mass market on the customer side. The shape made sense when scale required headcount and customers wanted standardized products.

Both halves are eroding.

On the supply side, AI lets one person do work that used to need ten. A six-person team can ship what would have taken thirty in 2018. The bottom of the pyramid thins out first.

On the demand side, customers no longer want one-size-fits-all. They want products tuned to their workflow, their language, their odd constraint. Mass market becomes hyper-personalized market.

The new shape is a million tiny craftsmen, each shipping a product to a few thousand customers, often across borders, often charging small monthly amounts. A solo founder serving 2,000 customers at $19 a month has a $456,000-per-year business with no boss and no headcount.

Big companies will not vanish. But the gravity of work has moved away from them. The center of mass is now in the long tail of small operators, and the AI tooling assumes that shape. Pyramids are losing tenants.

What the data says we should do better

The same report that flagged my strengths flagged my failure modes. I will not skip them, because if you are a power user, yours probably look similar.

Three things stand out. First, completion claims declared too early — the AI says "done" before runtime checks confirm it. The fix is to refuse "done" without evidence, every time. I am about half-disciplined on this.

Second, output token limits hit on long autonomous runs, causing truncation. The fix is to break work into smaller, checkpointed tasks rather than asking for one heroic run. I keep doing the heroic run anyway.

Third, ambiguous scope on creative work — blog drafts, rebrand plans — invites over-engineering. The AI fills the gap with what it thinks you want, which is usually too much. The fix is to lock the angle before writing the first prompt.

Look at your dissatisfied turns. The pattern is almost never "the AI is bad." It is almost always "I gave it the wrong shape of problem."

Summary

• Build is cheap. Intent and verification are the new craft.
• Juniors who iterate without ego often beat seniors who fight the tool.
• The CTO job is shifting from resource management to intent translation with evidence-based review.
• Solo founders will multiply, but distribution and domain depth still gate revenue. AEO is the new SEO.
• Pyramids are losing tenants. The center of work is moving to the long tail of one-to-few-thousand craftsmen.

None of this is settled. The shape of work in 2027 will still surprise us. But the way to find out is to build, ship, push back when the AI is shallow, and keep asking what kind of craft survives. That is the conversation I want to be in. Tell me what you see from where you sit.

FAQ

Do you need to know how to code to vibe code?

No, but you do need precise intent and the will to verify what the AI gives you. The hard part of vibe coding is not typing code; it is deciding what should exist, then pushing back when the output is shallow. People who already understand a domain — sales, design, logistics, law — can ship working products without writing every line.

Won't AI replace senior engineers eventually?

It is changing what senior means. Old senior was about knowing how to build; new senior is about knowing what to build, how to verify it, and where the load-bearing decisions are. Juniors who iterate without ego often outperform seniors who fight the tool because they have something to defend.

How do you measure a vibe coder's productivity?

Lines and commits are weak proxies. Better: number of intent-to-shipped cycles per week, defect rate caught by your own pushback, and how often the output runs in production without rework. In 29 days I shipped 275 commits and 238,530 lines, but the metric I trust most is that 23 real defects were caught the first time I said "verify this without guessing."

What is AEO and why should solo founders care?

AEO is Answer Engine Optimization — getting your product surfaced when an AI answers a user query. As more buying decisions start with "ask the AI which X is best," the moat shifts from SEO rankings to whether AI agents recommend you. Build is cheap now; distribution and trust are the part AI cannot do for you.

Related reading:

• How I made the AI recommend my app
• Harness engineering for vibe coding
• Spec-driven coding with Claude Code