LLM Debugging: Error Logs and Review LLM Debugging: Error Logs and Review | Zhiwei Li

LLM Debugging: Error Logs and Review | Generated by AI

Home 2026.03

Question: When debugging with an LLM, is it especially effective to copy browser console logs or HTTP request errors, and while some say you don’t need to read AI-generated code, isn’t being able to quickly review it still a bonus for getting better help?

Answer:

Debugging with LLMs: Console Logs, HTTP Errors & Code Review

Why Console Logs and HTTP Errors Are Gold for LLMs

Yes — pasting raw browser console output or network errors is one of the most efficient ways to get useful help from an LLM, for several reasons:

1. Errors are already structured information

  • Stack traces, error codes, and HTTP status messages are precise and unambiguous
  • LLMs are trained on massive amounts of similar error patterns from Stack Overflow, GitHub issues, MDN, etc.
  • There’s no “translation loss” — you’re giving the AI exactly what the runtime produced

2. HTTP errors are especially diagnostic

  • A 401 Unauthorized vs 403 Forbidden vs CORS error each tells a completely different story
  • Network tab details (request headers, response body, URL, method) give the LLM full context to pinpoint root cause
  • The AI can immediately distinguish “your token is wrong” vs “your server isn’t allowing cross-origin requests”

3. Console logs remove ambiguity

  • Instead of you describing a bug (“it doesn’t work”), the log shows the bug
  • Null reference errors, type mismatches, undefined variables — all self-evident in a log

The “You Don’t Need to Read AI Code” Debate

Some people argue you can just blindly run AI-generated code. That’s partially true for simple scripts — but quick review is genuinely valuable, and here’s why:

Scenario Skip Review Quick Review
Simple 5-line utility ✅ Fine ✅ Also fine
Code touching auth/security ❌ Risky ✅ Essential
Code modifying database/files ❌ Risky ✅ Essential
Large multi-function output ❌ Hard to debug later ✅ Catches mismatches early
You need to maintain it later ❌ You’ll be lost ✅ You’ll understand it

Even a 30-second skim helps you:

  • Spot if the AI misunderstood your intent
  • Notice hardcoded values that should be variables
  • Catch obvious logic errors before running
  • Give better follow-up feedback (“you used POST but I need GET”)

The Feedback Loop: The Real Power

The real synergy is the cycle:

You paste error → AI explains & fixes → You skim the fix
→ You run it → New error appears → You paste that error
→ AI narrows down faster because context is building

Each paste of a real error grounds the conversation in reality rather than speculation. The AI isn’t guessing what might be wrong — it’s reading what is wrong.

Practical Tips

  • Paste the full error, not a summary — details like line numbers and stack frames matter
  • Include the relevant code block alongside the error for even faster diagnosis
  • Network tab > Console for API issues — copy the full request/response
  • A quick scan of AI code is a low-cost, high-reward habit — even 20 seconds of review pays off
Back Donate