2 AM. AI is talking nonsense. Suggests code that doesn't compile. Forgets what it said a minute ago.

Thinking: is the model broken? Limits? Bug?

No. I'm tired.

Pattern

At 10 AM: clear prompt, specific task, good result.

At 2 AM: vague prompt, three tasks in one message, garbage output.

Garbage in

// Bad (2 AM)
make it work and also add that thing and fix the bug

// Good (10 AM)
Function validateEmail should:
1. Accept string
2. Return boolean
3. Check format with regex

AI reflects prompt quality. Clear prompt - clear answer.

What I do

When AI starts acting dumb - I don't blame the model. I look at myself.

• Tired? Go to sleep.
• Prompt is vague? Rewrite it.
• Too much in one message? Break it up.

Rule

If AI outputs garbage 3 times in a row - stop. Either I'm tired, or the task is poorly formulated.

Either way, the problem isn't the model.