Diagnosing bad outputs
The Four Diagnostic Questions
Every bad output is bad in a specific way. The four questions in this lesson are the only ones you need to figure out which way. Run them in order. The first one to give you a "yes, that's the problem" is your diagnosis — and your follow-up prompt writes itself from there.
The four questions
| # | Question | If the answer is "yes" |
|---|---|---|
| 1 | Is the format wrong? | Output spec was thin. Patch the output spec. |
| 2 | Is the length wrong? | Length wasn't pinned. Add a length constraint. |
| 3 | Is the content wrong? | Context was thin or facts were missing. Add the missing context. |
| 4 | Is the tone or voice wrong? | Role/persona was thin or absent. Patch the role slot. |
These map directly back to the five-slot anatomy from Module 2. Bad output → which slot was thin? Patch that slot. Send.
Worked diagnoses
Here's what each diagnosis looks like in practice.
Format
The model gave me a long flowing paragraph; I wanted a numbered list.
Diagnosis: format. Follow-up:
Same content, but reformat as a numbered list with each item on its own line.
Length
The model gave me 350 words; my Slack channel needs 100.
Diagnosis: length. Follow-up:
Cut to 100 words. Keep the conclusion; drop the background paragraph.
Content
The summary doesn't mention the budget freeze, which is the most important thing.
Diagnosis: content (you forgot to put it in context). Follow-up:
Important context I forgot to include: there's a hiring/budget freeze
through Q3. Update the summary to lead with that.
Tone or voice
The reply sounds corporate and stiff; we want warm and direct.
Diagnosis: tone. Follow-up (or update system prompt for next time):
Rewrite this in a warmer, more direct voice. One short sentence per idea.
Drop the "thank you for your feedback" framing.
Combining diagnoses
Sometimes more than one is wrong. That's fine. State all of them in one follow-up.
Same content, but: cut to 100 words, reformat as bullet points,
and shift the tone from corporate to plainspoken.
What you don't want is to fix one thing at a time over five rounds. The model loses track of what's stable and what's changing. State the corrections you want all at once, in one follow-up.
What you almost never need
Notice what's not on the list:
- "The model is dumb." (No, the prompt was thin.)
- "Try a different model." (Sometimes useful, but not the first move.)
- "Add 'be more creative' to the prompt." (This rarely fixes anything specific.)
- "Tell the model it did badly." (No effect on the next output. Skip the scolding.)
When you find yourself blaming the model, look back at the prompt instead. Nine times out of ten, one of the four diagnostic questions points at a slot you left empty.
A small habit
Print or save the four questions somewhere visible. For your first month of practising iteration, run them out loud whenever an output disappoints you. After a month they're internalised — you won't need the list — but for now the list is the muscle memory.
Next: a worked repair — the LinkedIn rewrite, fixed in three diagnostic-question moves. :::
Sign in to rate