Stop Vibe Coding (Do This Instead)

Overview

The subtitles explain why “vibe coding”—prompting an AI to generate code from vague intent—often works at first but then breaks down. They argue for a more reliable approach called agentic engineering.

---

Main Technological Concepts & Claims

“Ceiling of vibe coding” / compounding failure

• Early AI-generated code can look workable, but repeated prompting without understanding or structure introduces unknown additions.
• This compounds over time, making later changes harder and riskier.
• The resulting failure modes can include:
  • Apps breaking
  • Feature conflicts
  • Projects that freeze after starting fast
• The problem is framed as a process/definition issue, not luck or tooling:
  • The builder isn’t specifying what “correct” means upfront.

Shift from “vibe coding” to “agentic engineering”

• The video claims Andre Karpathy coined “vibe coding” in early 2025 and later suggested agentic engineering as a preferred term.
• Definitions:
  • Agentic: you generally aren’t writing code directly; you direct AI agents that do the work.
  • Engineering: treated as a skill you can develop, with real depth—especially the ability to direct work.

Director vs. prompter (same tools, different outcomes)

• A prompter gives broad instructions, such as: “Build me a user dashboard.”
• A director defines requirements before execution, for example:
  • which metrics to display
  • which data source to use
  • update frequency (e.g., every 30 seconds)
  • loading behavior while fetching
• Core distinction: clarity of intent before the AI produces anything.

---

Cited Reviews / Evidence (decline in trust)

• The video references a Stack Overflow survey of 49,000 developers:
  • Trust in AI-generated code declines from 40% to 29% over one year.
• The implication is that this matches real-world experiences:
  • code may appear correct superficially, but later fails in:
    • robustness
    • integration
    • maintainability

---

Guides / Tutorial: “3 habits” to Replace Vibe Coding

The video proposes a method that doesn’t require extra tools or coding expertise:

1. Define the output before opening the AI tool

   • Emphasize outcome/success, not just features.
   • If success can be described in two sentences, prompts become more precise.

2. Give the AI one job at a time

   • Avoid a single “build everything” prompt.
   • Break work into verifiable steps, e.g.:
     • build the login flow
     • build the dashboard
     • connect them

3. Review/test before prompting the next step

   • Verify what was just built to prevent “unknown breakages” later.
   • The claim is that this eliminates most cases of “it broke and I don’t know where.”

---

Product / Startup Analysis Mentioned

• The video claims that about a quarter of startups in the Y Combinator 2025 batch had codebases that were 95% AI-generated.
• It frames this as evidence that vibe coding accelerates initial shipping (e.g., days vs. months), but warns that without agentic-style definition and directing, it eventually hits the ceiling.

---

Main Sources / Speakers

• Andre Karpathy Credited with coining “vibe coding” and later proposing “agentic engineering” as a preferred term.

• Stack Overflow survey Cited as reporting a trust decline in AI-generated code (49,000 developers).