AIが行き詰まったとき

It happens to everyone: you paste an error, the AI suggests a fix, it doesn't work, you paste the new error, it suggests another fix... and round and round you go. The AI is now going in circles.

This isn't a dead end. There are reliable ways to break the loop and get unstuck. This lesson gives you the moves the pros use when Cursor, Claude Code, or Copilot lose the thread.

Understanding the cause helps you fix it. AI usually loops because:

• It's missing context — it can't see a file or value that matters
• It's guessing instead of seeing the real error
• The conversation got polluted with failed attempts it now clings to
• The real bug is somewhere it hasn't looked

Each of these has a counter-move. Let's go through them.

The #1 fix: start a new conversation. Once a chat is full of failed attempts, the AI keeps anchoring to its earlier wrong ideas.

Open a clean chat and re-explain the problem from scratch — but this time with everything you've learned. A fresh context window often solves in one shot what 20 messages couldn't.

If you keep the same chat, at least list the failed attempts explicitly. Otherwise the AI may suggest the exact thing that already failed.

This single sentence redirects the AI's search away from dead ends and toward fresh ideas.

Often the AI is stuck because it literally can't see the problem. Ask yourself: what does it not know?

• The actual value of a variable at runtime (log it and paste it)
• A related file it never saw
• The library version you're on
• What the data actually looks like

Paste those facts. Many 'stuck' loops are really 'starved for context' loops.

When AI loops, it's usually patching symptoms. Force it to stop guessing and investigate first.

This flips the AI from 'try random fixes' mode into 'diagnose like an engineer' mode.

If a bug is buried in a big messy file, recreate it in the smallest possible standalone snippet. Often, building that minimal example reveals the bug yourself.

Run this minimal reproduction of a classic async timing bug — it's small enough that the cause becomes obvious. Small examples make both you and the AI smarter.

Different AI models have different strengths. If one tool is stuck, paste the same problem into another — try Claude, then ChatGPT, then a different model in Cursor.

A fresh model with no baggage from your failed chat will often spot what the first one missed. There's no loyalty in debugging — use whatever gets you unstuck.

Sometimes you're both stuck because the assumed cause is simply wrong. Step back and challenge the framing.

Ask: 'Are we even sure the bug is in this function? What if the data coming IN is already broken?' Pointing the AI upstream — to where the bad value originates — frequently breaks the loop instantly.

For weird library or framework errors, the fix may not be reasoning at all — it may be a known issue. Copy the exact error into a web search, or check the tool's docs and GitHub issues.

You can even ask AI: 'Is this a known issue with [library] version X?' Sometimes the answer is 'upgrade the package' or 'a config flag changed' — something no amount of code-staring would reveal.

Keep this list handy. When you and the AI hit a wall, run down it:

• Fresh chat with current code + what failed
• List what you already tried
• Add missing context (logged values, related files, versions)
• Diagnose before patching — ask for ranked causes
• Shrink to a minimal example
• Second opinion from a different model
• Question the premise — look upstream
• Search for known issues, then take a break

You finished the Debugging with AI course! You now know that being stuck is never the end. When AI loops, you can:

• Start fresh, list failed attempts, and add missing context
• Force it to diagnose before patching
• Shrink the problem, get a second opinion, and question the premise
• Search for known issues — and take a break

Across this course you learned to read errors, hand AI great context, narrow bugs down, and break loops. That's a complete debugging toolkit. Now go build — and fix — fearlessly.