I've been using Claude Code for a couple of days
claude-codeagentic-codingllm-codingdeveloper-workflowhacker-news
Abstraction: Hacker News debate on Claude Code coding experience
Key points:
- HN thread (Mar 2025) debating LLM coding tools; recurring "coder vs. programmer" framing — LLMs excel when acceptance criteria is "good enough," fall short for opinionated programmers who care about maintainability/artistry.
- Claude Code / Claude 3.7 criticized for over-engineered, brittle, overly complex solutions; several note 3.5 "came up short" but was easier to iterate on than 3.7's rewrites.
- Practical workflows shared: treat the LLM as a fast junior dev; use spec documents, Cursor Rules / CLAUDE.md, a "stdlib" of rules (ref ghuntley.com/stdlib); limit context (~25k tokens), small files, separate dev/debug git branches.
- Cost anecdotes with Aider + Sonnet: throwaway scripts under $2; one user claims recreating a claude-code clone in a day for ~$20 (2200 lines, half unit tests).
- Widely cited failure modes: hallucinated APIs/validations (~20% made-up APIs), yes-man agreement, poor 2nd/3rd-order reasoning; strongest use as semantic search / Google replacement and boilerplate.
Connections: Claude Code · Anthropic · Cursor · Aider · Agentic Coding · Prompt Engineering · Large Language Models