Four failures you'll recognize a thousand times

You've learned to guide conversations. but a well-guided conversation doesn't guarantee the output comes back clean. Outputs fail with repeated patterns. If you can spot them at a glance, you stop "twisting the prompt" and go directly at the symptom.

On the right you have four cases. Each shows a reasonable prompt and an output that goes wrong. Before revealing the diagnosis, look at the output and try to name WHAT failed: did it invent? did it ignore the format? did it ramble? did it obey something that lived inside the data?

These four cases cover ~80% of the failures you'll see in production. Recognizing them quickly is the difference between "the model doesn't understand" and "I already know what to tweak."

When you've revealed the four diagnoses, move on to the close.