Field Notes from Shipping Real Code with Claude: Summary and Structured Insights

1. Introduction: Why Share Real-World AI Coding Tips

• The author notes that while there's no shortage of writing about LLM-based coding tools like Claude, they wanted to share genuinely practical tips learned in the field.
• Key quote

  "Honestly, there's a flood of Claude Code content out there right now. But we've found a few genuinely useful things worth sharing."

2. Core Tip: The Anchor Comments System

• What are Anchor Comments? A practice of leaving specially formatted comments throughout the codebase so they can be easily grepped later, recording context and important notes inline.

• CLAUDE.md example

  ### Anchor comments
  
  Add specially formatted comments throughout the codebase, where appropriate, for yourself as inline knowledge that can be easily `grep`ped for.
  
  - Use `AIDEV-NOTE:`, `AIDEV-TODO:`, or `AIDEV-QUESTION:` as prefix as appropriate.
  - _Important:_ Before scanning files, always first try to grep for existing `AIDEV-…`.
  - Update relevant anchors, after finishing any task.
  - Make sure to add relevant anchor comments, whenever a file or piece of code is:
    - too complex, or
    - very important, or
    - could have a bug
  

• Key keywords:

  • AIDEV-NOTE:, AIDEV-TODO:, AIDEV-QUESTION:
  • CLAUDE.md file
  • Add comments based on code complexity, importance, and bug potential

• Key quote

  "Add specially formatted comments throughout the codebase in appropriate places. Whenever something is too complex, very important, or could have a bug, make sure to leave an anchor comment."

3. Real-World Experience and Community Reactions

• Feedback from experienced engineers

  "I don't use LLMs systematically, but seeing how you actually use them on real projects is motivating. Thanks for sharing such a practical system."

• Comparison with other tools
  • aider vs. Claude Code: aider excels at memory/context management, but the author prefers Claude Code's TUI (terminal UI).

    "aider is a completely different tool. It's tops for memory/context management, but I prefer Claude Code's TUI. It ultimately comes down to personal preference and workflow."

4. Test Code and the Role of LLMs

• Humans should write tests!
  • The author strongly emphasizes prohibiting AI from writing or modifying test code.

    "Do. Not. Let. AI. Write. Your. Tests."

• Why
  • When AI writes tests, it becomes hard for humans to understand or modify them later, and developers grow complacent about testing — leading to more production bugs.

    "When someone later tried to modify AI-generated tests, it got really messy. And developers got so lazy about testing that bugs actually increased."

• How to enforce it
  • Specify in CLAUDE.md that test directories are off-limits
  • Disable editing tools for test directories in .claude/settings.json
  • Verify via PR reviews and commit messages that AI hasn't touched tests (largely trust-based)
• Counterargument
  • Some argue "having AI write a first draft of tests, then having humans review them, is also useful"

    "Having AI write tests first and humans review them carefully has been really helpful for me. The key is that the human takes final responsibility."

5. Choosing LLM Tools and Costs

• Claude Code and Opus 4
  • Opus 4 is the most capable model; the Max subscription ($100/month or $200/month) is recommended.
  • Some find token counting complicated.

    "The simplest way to start is to subscribe to Claude Max at the $100 tier and use Opus 4. Other models don't give you the real experience."

• Realistic cost considerations
  • It's possible to spend over $10 a day, with annual costs potentially reaching $2,000.

    "I tried Claude Code on a personal project a few days ago — it was incredibly efficient. But it's really expensive. I spent over $10 in a single day. Feels like paying an AI tax, but I can't help it."

• Offshore developers vs. LLMs
  • As LLM performance rapidly improves and costs fall, some companies are choosing LLMs over cheaper overseas developers.

    "LLMs are faster and cheaper than developers from lower-cost countries, so we've stopped hiring from those markets."

6. Organizational and Cultural Changes from LLM Adoption

• Transparency and team rules
  • Using LLMs doesn't make team rules disappear. Code reviews, linters, and team norms remain as important as ever.

    "Using Claude or other automation tools doesn't mean team rules go away. Upholding the team's standards through code review and linters is still crucial."

• AI discomfort as a signal to systematize
  • Where AI feels uncomfortable to use can be a signal that the area needs a proper validation system.

    "If something feels uncomfortable to hand off to AI, that might be a signal that you need to introduce a systematic verification process for that area."

• The role of documentation and comments
  • CLAUDE.md, SPEC.md, AIDEV comments, etc. must provide clear rules and examples so the LLM understands.
  • Documentation for humans and documentation for LLMs differ in style and content.

    "A style guide for humans might be 100 lines, but a guide for Claude needs over 500. You have to include far more examples."

7. Real Workflows and Productivity

• Large-scale refactoring with LLMs
  • Refactored 500+ endpoints in 4 hours (excluding tests)

    "Tests weren't included — that took much longer. Now our developers can't use 'I wrote the tests sloppily' as an excuse."

• Integrating documents and research
  • Claude Code's agentic markdown editing enables semantic merging of multiple sources into one document

    "When you paste it into Claude Code, instead of a simple merge it can do 'integrate the relevant parts of this research into this document' — a semantic merge. It's incredibly powerful!"

8. Community Debate: LLMs and Authenticity

• The transparency controversy
  • It became controversial that 40–60% of the post was LLM-assisted.
  • HN moderators saw a split between "content quality is what matters" and "authenticity matters more."

    "If the most interesting and insightful content came from an LLM, isn't that what we should be reading and discussing?" "No. I'd rather read something boring that a real human actually wrote."

• The author's response
  • The author clarified: "The ideas, code, examples, and images are all mine; the LLM helped with drafting and research."
  • Moderators restored the post to the front page, noting: "We can allow non-native English speakers to use LLMs to polish their writing."

9. Conclusion: The Future of LLM Tools and the Developer's Role

• LLMs aren't perfect yet, but they're improving fast
  • "A new lever is being built. It's still rough, but worth learning."
• Developer responsibility
  • LLMs speed up coding, but testing, verification, and final accountability remain with the developer.

    "As feature development gets faster, having developers write their own tests keeps them accountable for the code and its bugs."

Key Takeaways

• Introducing systematic collaboration methods with LLMs — such as Anchor Comments and CLAUDE.md — can improve both code quality and productivity.
• It's best for humans to write test code directly; even when LLMs help, final verification and responsibility belong to the developer.
• Costs are non-trivial, but LLM performance and efficiency are advancing rapidly.
• Team culture and rules remain important after adopting LLMs; transparency and accountability must be maintained.
• There are debates about AI-assisted authenticity and content quality, but what ultimately matters is real value and results.

💡 After reading this, you'll understand not just how to use LLMs, but why to use them this way — and what real-world struggles and lessons learned in the field look like.

Key keywords:

• Anchor Comments
• CLAUDE.md
• AIDEV-NOTE, AIDEV-TODO, AIDEV-QUESTION
• Human-written test code
• LLM tools (Claude Code, Opus 4, aider, Cursor)
• Cost / subscriptions
• Team rules, code review
• Transparency, authenticity, accountability

😊 Feel free to ask if you have any questions