Signing in

You will be sent to MillerKnoll sign-in.

AI

Debug a Bad Run

A bad output is data. Before you give up on AI for a task, ask which layer failed: the prompt (missing goal, context, or format), the source (wrong or incomplete material), or the expectation (you asked for finished work when you needed a first draft).

Lesson 4

Prompt, source, or expectation?

This is the same diagnostic instinct as debugging machine learning models — simplified for knowledge work. Change one layer; rerun; observe.

If three focused fixes still fail, the lesson may be that this task needs human judgment, better source data, or an agent with pinned knowledge — not more prompt tweaking.

Core principles

  1. Prompt failure: vague goal, missing audience, no format — fix with GCSE or Role+Task+Format.
  2. Source failure: Copilot never saw the file, transcript, or data you assumed — attach or paste source explicitly.
  3. Expectation failure: you needed verification-heavy facts but treated output as final — adjust tier and task type.
  4. Fix one layer per rerun. If you change everything, you learn nothing.
  5. After eight refinement rounds on the same prompt, rewrite the original brief — not another tweak.

Go deeper: Prompt Engineering — why prompts fail

Check yourself

What should you change between debug reruns?

Do this in Copilot

Take your weaker output from Lesson 3. Label the failure mode. Fix one layer and rerun once.

Paste this into Copilot Chat and work through it before moving on.

Diagnose a weak output

Here is a Copilot output that missed the mark for [TASK]: [PASTE OUTPUT]. My prompt was: [PASTE PROMPT]. Diagnose whether the main failure is prompt, source, or expectation. Recommend one specific fix — not a full rewrite.
Open Copilot →
  • Failure-mode diagnosis

Did you run this in Copilot? Mark complete when you have tried it.

Next lesson: Report What You Learned →