Prompt Debugging · ai-skills

Asking Clearly · Beginner to Intermediate

4topics
4lessons

Most bad replies trace to the prompt rather than the model. Debugging a prompt means classifying the failure, changing one thing, and rechecking a small fixed set of test prompts.

Symptom	Likely class	First test
Reply invents facts the prompt never gave	Missing context	Add the missing facts and rerun.
Reply follows a different reading than intended	Ambiguity	Define each key term once and rerun.
Right content in an unusable shape	Wrong format	Pin the shape with one example.
Failure survives every phrasing	Capability limit	Split the task or check the step another way.

The prompt debugging loop.

Each red reply is classified, one element changes, and the full set decides.

Lesson	What it covers
Diagnose the failure first	Four failure classes and the test that tells them apart.
Change one thing at a time	Single-variable edits with a before-and-after log.
Keep a small set of test prompts	Test prompts with checkable expected properties.
A prompt regression harness	A runnable Python harness over a scripted fake model.

Check yourself A reply is correct but arrives as prose when the flow needs a table. Which class is this, and what is the fix?

Wrong format. Pin the shape in the prompt, for example by naming the columns once.

Up nextDiagnose the Failure First

Practice this lesson Report an error