Claude Code as a Daily Driver: Claude.md, Skills, Subagents, Plugins, and MCPs

Original Article ↗ Hacker News Discussion ↗

Power users describe elaborate workflows built around Claude Code—using files like CLAUDE.md, custom “skills,” subagents, and tools such as Nix—to let the model autonomously modify large codebases, run tests, and even write commits. Many report major productivity gains and see Claude as a genuine force multiplier, but others criticize the complexity, vendor lock‑in, slowness, and tendency to produce opaque or low‑quality code that’s hard to maintain without the same agentic setup. A recurring theme is the tension between treating LLMs as powerful but fallible tools that need guardrails, versus outsourcing too much of the development process to an always‑online, proprietary system.

Costs and Usage Patterns

  • Reported costs range from ~€10–22/month for Claude Code via Pro or API, up to $100/month for heavier Opus use with care taken to manage context length and rate limits.
  • Some users say Claude Code saves significant debugging and boilerplate time; others feel it just generates more work or low‑value bug reports.

Workflows, Files, and Automation

  • Many rely on project config files like CLAUDE.md and VOCABULARY.md to define conventions, terminology, commit message style, and desired behaviors.
  • Pre-commit hooks and deterministic scripts are used so agents must pass tests/linting instead of “remembering” to run them.
  • Some prefer lightweight use (“just an IDE with Claude integration” or simply prompting it to run existing CLIs) over elaborate skills/subagents/MCP setups.

Environment Management (Nix, Docker, etc.)

  • Nix integration is praised for reproducible dev/test/prod environments and sandboxed agent work; others are happy with Docker or even a dedicated VPS.
  • Alternative tools (Mise, uv, local VMs, etc.) are mentioned as simpler or more familiar trade‑offs.

Reliability, Downtime, and Vendor Lock‑In

  • Strong concern about depending on an always‑online model for core SDLC work; comparisons are made to CAD, version control, and other cloud tools.
  • Some argue you can just swap to another harness/model (Codex, DeepSeek, OpenCode, local models); others say prompts and skills are highly model‑specific, leading to implicit lock‑in.
  • Worries include: inability to take over an LLM‑written codebase when the model is down, sudden price hikes, and long‑term dependence.

Quality, Autonomy, and “Slop”

  • Several users find agentic workflows powerful for large codebases, as long as they retain human review and limit autonomy.
  • Others report cut corners, shallow tests, hallucinations, and ignored instructions, leading to distrust and frustration.
  • There is disagreement over delegation vs. tight, stepwise guidance: some see “delegate, don’t pair‑program” as efficient; others say it produces opaque, hard‑to‑maintain “slop” codebases.

Skills, Commands, and Complexity

  • Many view skills/commands/subagents/plugins as mostly “canned prompts” with overlapping purposes and confusing redundancy.
  • Some argue they add unnecessary accidental complexity; others see value in standardized review flows (e.g., structured /code-review with effort levels) but still question token cost and actual bug‑finding effectiveness.

Cultural Backlash and Content Quality

  • Multiple commenters complain about AI‑generated, repetitive “how to use coding agents” posts, calling the ecosystem hype‑driven and cultish.
  • Skeptics dislike having to hand‑craft elaborate scaffolding to make “smart” tools usable, and worry that responsibility for failures is being pushed onto users rather than vendors.