Claude Code 2.0

Original Article ↗ Hacker News Discussion ↗

Claude Code vs Other Coding Agents

  • Several users compare Claude Code with Goose, Aider, Codex, OpenCode, Cline, Crush, etc.
  • Common praise: Claude Code’s orchestration (planning, tools, hooks, shell integration) and CLI UX feel more polished and “native” than most competitors.
  • Goose can be made to behave similarly but often needs heavier manual prompting; Claude’s default system prompt and planning are seen as a big differentiator.
  • Some prefer open-source/any-model tools (OpenCode, Aider, Cline) for provider flexibility, SDKs, and LSP features, even if they’re buggier or less polished.

What’s New in 2.0 (and Mixed Reactions)

  • Noted changes: checkpoints + /rewind (rewind context and code edits), /usage limits view, plan/auto-accept toggling via Shift+Tab, Tab to toggle “thinking,” history search (Ctrl‑R), VS Code extension, and altered prompt.
  • Checkpoints are widely welcomed as a “git-like undo” tied to conversation context, though power users still rely on git/Jujutsu for more flexible history.
  • Some regressions/annoyances: loss of inline filename tab-complete, TUI now starting full-screen, VS Code plugin missing CLI features, plan UI changes, CJK input issues, and high RAM use from the Node/React stack.
  • Questions and confusion around “Plan with Opus, implement with Sonnet” removal; workaround model options (/model opusplan, etc.) are discussed.

Prompt, Comments, and Planning Behavior

  • Extracted system prompts show more explicit task lists, XML-like tags, and changed guidance (e.g., comments, emojis).
  • Large side-thread on whether auto-inserted comments are helpful or “instant technical debt,” with strong opinions both for and against.
  • Debate over Sonnet 4.5 vs Opus 4.1 for planning; benchmarks vs lived experience, especially on hard reasoning tasks.

Effectiveness and Limits

  • Users praise Claude Code for boilerplate, refactors, restructuring, CRUD endpoints, and tedious wiring; often “fun again” for burned-out developers.
  • Weak spots: deep UX/api design, obscure bugs, complex networking/audio/edge cases, and large-but-simple refactors that exhaust context or cost.
  • Many adopt workflows with TODO.md/Kanban.md, planning docs, and explicit instructions to keep the agent on track.

Security, Safety, and Data

  • Strong disagreement on risk: some run in “YOLO mode” and see no problems; others report real incidents (wrong kubectl patches, config deletion, etc.).
  • Repeated warnings about prompt injection via dependencies, docs, PDFs, or curl, and calls for containers, restricted users, bubblewrap, Nix shells, or command allowlists.
  • Concern over Anthropic logging usage/conversations even with “training opt-out”; confusion over what exactly is stored and for how long.

Beyond Coding

  • Many use Claude Code on arbitrary folders: writing, research, requirements Q&A, music mastering scripts, video processing, D&D prep, reverse engineering, admin tasks.
  • Some argue “coding agent” is really an early general computer agent that can do anything a human with a shell can, raising both excitement and safety concerns.