Chapter C08 · Interactive Course
Improve one thing at a time
Diagnose a disappointing result and revise it on purpose.
Core Learning Concept
Repair the response in a controlled sequence. Identify the unmet requirement, ask for a specific change while protecting correct details, and compare the revision with the brief. Stop when it meets the need, or switch methods when repeated attempts are not resolving the problem.
What You’ll Learn in This Chapter
-
Find what went wrong
Identify the mismatch between a request and its result.
A mismatch is the difference between what you requested and what came back. Check the result against the brief: is the audience wrong, a fact missing, or the format unusable? Naming the failure gives you a focused repair and helps preserve parts that already work.
-
Change one useful ingredient
Make a targeted edit that can be evaluated.
Changing one ingredient makes a comparison easier to interpret. If the tone is wrong, revise the tone instruction while holding the facts and format steady. AI outputs can vary anyway, so one improved response does not prove causation. Repeat a representative check when the decision matters.
-
Clarify before assuming
Resolve missing information instead of inviting an invented assumption.
An unresolved detail is an unknown, not an invitation to invent. “Tomorrow at seven” may leave the date, timezone, or morning versus evening unclear. Ask the person who knows, or require the output to mark the detail as missing. Confirm it before sharing a definitive announcement.
-
Give specific feedback
Request a revision by naming what to keep and what to change.
Useful feedback names the change and what must stay. “Shorten the opening to one sentence; keep the confirmed date and venue” gives a revision target. After the edit, check both requirements. An instruction to preserve a fact guides the model but does not guarantee that the fact survives.
-
Split a large task
Break a compound request into inspectable steps.
A large task becomes easier to supervise when intermediate results can be inspected. For a newsletter, agree on an outline before drafting every section. Each checkpoint should answer a real question, such as whether the topics and order fit; dividing work into arbitrary tiny steps adds effort without useful control.
-
Know when to stop
Stop iterating when the output meets the need or a different approach is warranted.
A stopping rule tells you when another round is no longer worthwhile. Stop when the result meets your criteria, when the time budget is reached, or when another method is needed. A different wording is not automatically an improvement. Compare progress with the actual purpose of the task.
-
Repair a disappointing response
Diagnose a missing requirement, make a targeted revision, and decide whether the next draft is ready or needs another kind of help.
In-Depth Practice & Real-World Application
Workplace Scenario: Diagnosing an unsatisfactory AI answer and steering it toward perfection systematically.
Worked Example: One-Variable-at-a-Time Prompt Refinement
Common mistake: Rewriting the entire prompt from scratch when the tone is too formal and a key section is missing.
Effective pattern: Keep the prompt, but adjust one knob: “The content is correct, but change the tone from formal academic to conversational peer-to-peer. Keep all bullet points unchanged.”
Changing everything at once makes it impossible to know what improved the result. Iterating one variable isolates what works.
Key Takeaways
- Never abandon a decent first draft: treat the first generation as raw clay to shape.
- Isolate variables: adjust tone, length, structure, or content independently.
- Use critique prompts: ask the AI “What assumptions did you make in this draft?” to find blind spots.
Workplace tip: When an answer falls short, don’t start over. Reply with: “Keep [part X], but rewrite [part Y] with [specific requirement Z].”