What Happened
Peter Bloem published an essay titled “AI Coding Without the Vibes” arguing that “vibe coding” — the laissez-faire approach where you prompt an AI and accept whatever it spits out — is the wrong model for professional work. His alternative: craft coding. You write the code. AI reviews it. You stay in the loop.
The piece hit Hacker News and sparked 55 comments (97 points) in two days. Not a flamewar — a genuine debate among working developers about how they actually use these tools.
The Core Tension: Vibes vs Craft
Bloem defines vibe coding as “deferring to the AI, not worrying about individual lines, keeping an eye on whether tests pass.” It’s the trillion-dollar bet: if professionals work this way, the industry wins. If they don’t, the bet fails.
Craft coding flips it. You write. AI reviews. You remain the architect.
The HN thread didn’t just agree or disagree. It mapped the spectrum of real workflows.
Developer Reactions: The Spectrum
Camp 1: AI as Understanding Tool (Not Generator)
Multiple commenters converged on the same insight: the best use of AI isn’t writing code — it’s explaining code.
“LLMs are great with focused questions, up and down abstraction layers and across all kinds of concerns. Stack up these focused concerns into a rich understanding of what you are doing or writing. Understanding is the real output, code is the byproduct.”
This appeared in multiple forms. One developer uses AI to ask: “What is this bit doing? Is there a risk of data leaking here? Are permissions enforced downstream?” They treat it as a senior engineer sitting beside them — not a code generator.
Another put it bluntly: “Code review can go much deeper now if you use AI to aggressively attack a PR combined with your human insight.”
Camp 2: The Deskilling Fear
A strong current runs through the thread: juniors never learn if they outsource thinking to the machine.
“Proficient developers are getting pressured into outsourcing their craft and skills to the machine, thus deskilling quickly. Juniors never get the chance to become proficient in the first place — either because they’re not getting hired to begin with, or because they succumb to the siren song of the machine.”
The calculator analogy came up repeatedly. In the 70s, students weren’t allowed calculators until 10th grade. You learn by heart so when the calculator lies, you feel it. AI removes that feedback loop entirely.
One commenter noted: “People simply plug in their question and don’t even bother reading the output slightly. It’d be like seeing a planet mass of x^-27 and not realizing you input something wrong because you simply trust the calculator.”
Camp 3: The “Surgical Edits” Workflow
Several experienced developers described nearly identical workflows:
- Design first — human thinks through architecture
- Small diffs — ask AI for tiny, specific changes
- Review everything — read every line before committing
- Adversarial review — have a second AI pass critique the result
“My current approach is ‘ask for very small diffs’ + ‘review them very carefully.’ The frontier models can’t reliably modify Pong without breaking it. If you leave any room for misunderstanding, they will laser in on it and do the stupidest thing possible.”
This isn’t vibe coding. It’s micro-management with a very fast typist.
Camp 4: Core vs Periphery
A practical heuristic emerged: draw a line between what you care about and what you don’t.
“Core modules: coded by myself, AI reviews and AI to discover/learn. Stuff I don’t care about: API layer, CLI layer, smoke tests, integration tests — dial AI heavy, lighter human reviews accordingly.”
This mirrors how teams already work: senior engineers own the kernel; juniors and automation handle the plumbing. AI just became a very capable junior.
Camp 5: The “Read-Only” Debate
Bloem argues AI should be read-only — no codebase access, no running tools. Commenters pushed back.
“A good code reviewer will check out your PR branch, run your tests, and ideally try to break them. An agent can run your code in its own sandbox, try variations to see what breaks.”
The counterpoint: if you trust AI to review, why not trust it to run tests? The distinction between “reviewer” and “actor” blurs fast.
Camp 6: Multi-Model Verification
A surprising workflow: use one AI to check another AI’s work.
“I regularly use ChatGPT to check the code generated by Claude. I’ve recently settled on DeepSeek for one project and had it review other models’ output — that was quite eye-opening. Someone in the frontier-models part of the world is definitely paying attention to the AI slop generated by other models.”
This treats models as an ensemble. Each has different failure modes. Cross-checking catches what single-model review misses.
Camp 7: The XDG/Linux Config Side-Thread
A tangent became its own discussion: the Pi coding agent ignores XDG Base Directory spec on Linux. Config lives in ~/.pi instead of ~/.config/pi/.
“I’ve been seeing this pattern so frequently of obvious simple things like XDG being ignored. Cynically, I think it’s the result of inexperienced people suddenly being able to make tools. They never learned all these standards, so if an LLM spews out a junky solution, they’re none-the-wiser.”
This meta-point reinforces the deskilling argument: AI-generated tools inherit AI’s blindness to standards.
Real Insight: The Industry Is Splitting
The thread reveals three emerging developer archetypes:
| Archetype | Workflow | Risk |
|---|---|---|
| Vibe coder | Prompt → accept → ship | Deskilling, security holes, unmaintainable code |
| Craft coder | Write → AI reviews → iterate | Slower initial velocity, but sustainable |
| Hybrid | Core: craft. Periphery: vibe. | Requires discipline to maintain the boundary |
Most commenters say they’re craft or hybrid. But several admitted: “It takes a lot of patience and very easy to sin.”
The gravitational pull toward vibe coding is real. It’s faster. It feels magical. The discipline to stay in craft mode is exactly what the industry hasn’t proven it can sustain at scale.
What It Means for You
If You’re Learning to Code
Don’t outsource the struggle. The struggle is the learning. Use AI to explain concepts, not to write your assignments. One commenter’s rule: “You can use the electronic tuner once you learn to tune by ear.”
If You’re a Senior Engineer
Your value shifted. You’re no longer the best typist. You’re the architect, the reviewer, the person who knows when the AI is hallucinating a security vulnerability. Build workflows that force you to read every line.
If You’re Managing a Team
Define the boundary explicitly. “Core vs periphery” isn’t a vibe — it’s a policy. Document which modules require human authorship. Enforce it in CI.
If You’re Building Tools
Respect platform conventions. The Pi/XDG discussion isn’t bikeshedding. It’s a canary: tools built by people who never learned the basics will break in ways AI can’t explain.
The Uncomfortable Truth
Bloem’s essay ends with a challenge: vibe coding is the only model that justifies the trillion-dollar AI bet. If professionals don’t work that way, the economics collapse.
The HN thread suggests professionals aren’t working that way — not the good ones, anyway. They’re using AI as a force multiplier for understanding, not a replacement for judgment.
Whether the market rewards that distinction remains open. But the developers in this thread have already voted with their workflows.
Source: Hacker News discussion on “AI Coding Without the Vibes” — 55 comments, 97 points, August 2026.