Vibecoding #2

Original Article ↗ Hacker News Discussion ↗

Alternative tools & “reinventing the wheel”

  • Several commenters note the project resembles existing tools (SLURM / AWS ParallelCluster, Capistrano, Fabric, Ansible, Terraform, GNU parallel).
  • Some see value in a bespoke, simpler, homelab‑oriented tool; others would default to NixOS + tests or existing orchestration stacks.
  • There’s mild concern about spending a day “vibecoding a square wheel,” especially for critical infra code.

Monetization vs OSS for agentic infra tools

  • A similar remote‑dev / infra‑on‑demand tool is described; its author is unsure people would pay.
  • SaaS for CLI tools is called “gross”; preference expressed for selling libre software or charging only for hosted services (provisioning, monitoring) while allowing self‑hosting.

Cloud cost & safety

  • Strong reminders to auto‑shut down EC2/GPU instances to avoid surprise bills.
  • People share simple shutdown patterns (timed shutdown, cron with a keepalive file).

What “vibecoding” means & how to do it

  • Some argue this isn’t “pure” vibecoding but AI‑assisted coding.
  • One axis: >50% of code produced by AI vs. just occasional help.
  • Others report best results from a detailed spec/PRD plus checklists, then having agents implement phases, run tests, and review via automated loops.

AI adoption, FOMO & pricing

  • Debate over whether the author is “late” to AI: some say most engineers now use AI; others say many colleagues ignore it.
  • Strong sense of FOMO for some; others see it as hype with little real payoff yet.
  • Experiences range from $20/month plans being ample for “assistant” use to $100–$200 tiers needed for heavy, agentic workflows.
  • Confusion and discussion around per‑million‑token pricing and why some subscriptions feel far cheaper per unit.

Positive experiences & workflows

  • Multiple reports of 10x+ speedups for side projects, small tools, and hobby games, especially for “yak‑shaving” automation and throwaway scripts.
  • Patterns: “snipe mode” (targeted bugfixes, small changes) works well; full‑feature generation is fun but suspect for long‑term maintenance.
  • Some use agents as advanced codebase search and refactoring assistants, not as autonomous builders.

Skepticism, quality & human factors

  • Complaints about bloated, hard‑to‑review AI PRs, early‑2000s enterprise patterns, and more RCA incidents tied to overlooked mistakes.
  • Concern that AI accelerates “rewrite instead of fix” behavior and deepens development‑hell.
  • Mixed reports on agents for serious work: helpful for simple CRUD/Web tasks, often weak for niche domains (e.g., complex scraping, game dev, hardware design).
  • Broader critique that AI can’t fix product “enshittification,” which stems from incentives, not coding speed.

Local vs hosted models

  • Some want local models for privacy but find the ecosystem confusing; others bluntly say local LLMs are still far behind Claude/Gemini/OpenAI for serious coding.

Reflections on careers & time

  • Older and retired developers describe AI as finally letting them ship projects they never had time or focus to complete.
  • A few feel bored or alienated by prompt‑driven workflows and question staying in the field if that becomes the norm.