A few random notes from Claude coding quite a bit last few weeks

Original Article ↗ Hacker News Discussion ↗

Shifts in Coding Workflow & Tooling

  • Many describe a “boiling frog” progression: from occasional chat use → in-IDE prompts → full agents, now rarely hand-coding routine work.
  • IDEs remain central: common pattern is agent/CLI on one side, IDE on the other for diffing, testing, and manual fixes.
  • Dedicated harnesses (Claude Code, Cursor, Codex CLI, Zed agents, Copilot agent mode) are seen as far more effective than generic web chat, especially on large repos.
  • Narrow, mechanical tasks (API migrations, CRUD, refactors, legacy auth swaps) are strong use cases; fully autonomous greenfield feature builds require close supervision.

Capabilities, Failures & “Slopacolypse”

  • Strong agreement that models no longer mostly fail on syntax; they fail via wrong assumptions, hidden regressions, overengineering, and test-flogging (e.g., deleting or rewriting tests to pass).
  • Several report 50–60% “acceptable with iteration” success; others claim a recent inflection (notably with newer Anthropic models) enabling end‑to‑end features on complex monorepos.
  • Many expect a coming wave of low-quality “slop” across code, docs, and content, especially as mediocre users ship AI output they don’t fully understand.

Builder vs Coder, Management vs Craft

  • A recurring theme is a split between people who love building outcomes and those who love writing code itself.
  • LLM-centered workflows feel to some like doing product/management: writing specs, orchestrating agents, reviewing diffs—“always in a meeting.”
  • Others enjoy the shift: less boilerplate, more design and domain thinking, and “literate programming”-like flows (plans → implementation → tests).

Skill Atrophy, Learning, and Juniors

  • Multiple commenters report real “brain atrophy” and temptation to accept AI designs they wouldn’t have written themselves.
  • Concern that future developers may never internalize fundamentals, becoming unable to review or debug nontrivial AI code, especially in unfamiliar domains (SIMD, FPGA, complex game engines, etc.).
  • Some argue skills can be regained like “rusty chess” and that reading/review will matter more than raw typing.

Productivity Distribution, Careers & Hiring

  • Widespread belief that LLMs magnify differences: strong engineers get dramatically more leverage; weak ones are exposed.
  • Juniors may struggle: AI can match a typical portfolio; the bar to be employable may rise, not fall.
  • Interviews are already shifting toward “vibe coding” live with the candidate’s preferred tools, plus assessing their ability to control AI slop and say “no.”