After two years of vibecoding, I'm back to writing by hand [video]

Scope and Balance of AI-Assisted Coding

• Many commenters say current agents can’t replace hand-written code but are useful for tedious, low-risk tasks: small scripts, boilerplate, config wiring, refactors, and tests.
• Good fit: non-critical tooling, one-off utilities, CRUD frontends on top of robust backends, Streamlit/Shiny-style demo apps.
• Poor fit: critical systems (payments, ERP, core business logic), complex math/parallelism, architecture and design.

Vibecoding vs Targeted Assistance

• “Vibecoding” (letting agents build entire apps from vague specs) is widely seen as brittle: results may “work” but be incoherent, hard to maintain, or subtly wrong.
• Several people report that asking an LLM to complete an app from epic-level descriptions “kinda works” for toy projects, but is clearly unacceptable for real products.
• A recurring complaint: tools over-refactor, add unnecessary complexity, or touch many files when a simple, local fix would suffice.

Responsibility, Quality, and Technical Debt

• Strong consensus that responsibility for code remains entirely with the human; AI won’t be blamed when things fail.
• Concern that management will use AI to push “ship faster” culture, increasing volume of low-quality code and incidents.
• Some argue AI can actually improve rigor when used with strong artifacts (design docs, tests, structured agents); others see it mainly as a way to generate more technical debt faster.
• Tests and green CI are called out as a false sense of safety when coverage or assertions are weak.

Effects on Thinking and Craft

• Several note AI helps with “kinetic” coding (typing, boilerplate) but can weaken the developer’s mental model and architectural thinking if overused.
• Others argue it frees time to think more deeply about real problems, analogous to moving from low-level programming to higher abstractions.
• Some express discomfort or sadness at losing the “tasty bits” of hands-on problem solving and learning, especially when AI is used to implement things “too hard” for the programmer to understand.

Careers, Hiring, and Industry Dynamics

• Many aren’t personally worried about being replaced “right now,” but are worried about:
  • Perception-driven hiring freezes and expectations that fewer devs can do more with AI.
  • Especially grim prospects for juniors, who struggle to get initial experience.
  • Difficulty distinguishing real skill from LLM-boosted interview answers and contractor work.
• Self-driving cars are a popular analogy: big gains for assistance, but fully autonomous replacement may be much farther away than hype suggests.

Middle-Ground Practices

• Common advice:
  • Use AI for small, well-scoped tasks and boilerplate; avoid giving it end-to-end ownership.
  • Break work into small tickets, keep a refactoring backlog, and enforce code review and CI equally for human and AI changes.
  • Be explicit in prompts (e.g., no extra refactors, minimal changes) and treat the model like a junior dev whose work must be checked.
• There’s broad rejection of all-or-nothing positions: both “LLMs are useless” and “LLMs will do everything” are seen as unhelpful extremes.