What makes Claude Code so damn good

Original Article ↗ Hacker News Discussion ↗

Claude Code vs Other Tools

  • Many compare Claude Code favorably to Gemini CLI and GitHub Copilot Agents when used as a terminal/agentic tool: better tool use, instruction-following, and multi-step “terminal” workflows.
  • Gemini’s web UI is often praised, but its CLI is widely described as buggy, looping, overly verbose, or stuck in self-deprecating ruts; some nonetheless find Gemini better for whole‑repo understanding, systems programming, or architectural planning.
  • Opinions on Copilot (with Sonnet 4) and Cursor are mixed: some say Claude Code is clearly superior, others see little difference or prefer Cursor/Copilot for IDE integration and UX.
  • Some argue the “Claude is best at coding” narrative is partly marketing, especially if you’re not using the Claude Code CLI with repo context.

Productivity and Use Cases

  • Enthusiasts report dramatic productivity gains, treating Claude Code like “several junior devs,” especially for:
    • Greenfield projects, prototypes, MVP SaaS apps.
    • Boilerplate, tests, refactors, and unfamiliar stacks.
  • Others find only modest or negative gains, especially on:
    • Large, legacy, monolithic or niche stacks (C/C++, Rust, COBOL, proprietary systems).
    • Work where design, domain understanding, or novel algorithms are the main bottleneck.
  • Several note that perceived productivity can exceed measured productivity; style preferences and tolerance for “LLM-flavored” code heavily influence satisfaction.

Limitations and Failure Modes

  • Common complaints: circular editing, half-baked refactors, regressions in unrelated code, plateauing as projects grow, timeouts and very slow runs.
  • Specific pain areas: Elasticsearch, Security Onion, some localstack setups, Elixir/Phoenix without extra tooling, snarky or odd “personality” from some models.
  • Many emphasize you must still read and guide the code; fully autonomous “vibe coding” regularly creates technical debt.

Agent Design and Prompting Strategies

  • The article and thread highlight: long, explicit system prompts; flat tool loops instead of complex multi-agent routing; clear task lists; strong validation/feedback loops.
  • Simple heuristics (“focus on what it should do,” “THIS IS IMPORTANT,” “think harder/ultrathink”) are reported to materially change behavior, underscoring current steering limitations.
  • Some find custom sub-agents in Claude Code worse than the main agent, reinforcing a “keep it flat and simple” approach.

Tooling, Internals, and Security

  • Reverse‑engineering tools (e.g., trace/bridge utilities) help inspect Claude Code’s internal prompts and tool calls; the distributed JS is minified and not truly open source.
  • There’s interest in OSS alternatives (OpenHands, Cursor Agent CLI, others).
  • Security-conscious users worry about granting full CLI access; mitigations include per-command approval, directory sandboxing, Docker/VM isolation.

Broader Reflections and Skepticism

  • Debate over whether LLMs will homogenize stacks around “LLM-friendly” languages.
  • Ongoing tension between enthusiasts, skeptics, and concerns about hype, shilling, and lack of concrete public examples of successful, LLM-built products.