Study notes · 3.3% of the exam

Adjust approach based on feedback and results

Change the right thing in response to results: give specific feedback on a draft, add or diversify examples, add constraints, split the task, or switch feature or model when the prompt is not the problem.

Key points

  1. 1

    When a draft is close but off in tone, length or emphasis, reply with precise feedback naming what to keep and what to change; do not restart from scratch or accept the poor fit.

  2. 2

    Examples raise quality but the model copies them. If outputs become uniform (same structure, same offer for every case), diversify the example set and state what the examples illustrate (tone, structure) versus what varies.

  3. 3

    Removing all examples after over-fitting returns you to the weaker state; adding more of the same example makes over-copying worse.

  4. 4

    Add explicit constraints when outputs are too long, miss required elements or drift: required sections, maximum length, forbidden content. Constraints work better than adjectives.

  5. 5

    When an early step's errors (data cleaning, calculations) contaminate later steps (narrative, slides), decompose the work and insert a verification point after the step that matters; regenerating downstream content alone reproduces the errors.

  6. 6

    Recognise when the prompt is not the problem. Questions needing current, cited facts require web search or research mode; a bigger model, 'use the latest information', or format examples cannot supply information the model does not have.

  7. 7

    Switch model when the task's demands changed: a faster, lower-cost model for high-volume, straightforward work; a more capable model for deep reasoning. Switching models is not a fix for missing context.

  8. 8

    Iterate systematically: define success criteria before changing anything, test each version on the same representative inputs, change one element at a time, and keep a version log so you can return to a known-good prompt.

  9. 9

    Model self-ratings are not a quality measure; your own criteria and human review are. Separate personal copies of a shared prompt destroy consistency.

  10. 10

    Fixing one complaint and causing another on each revision is the signature of iterating without fixed criteria and a fixed test set.

Test yourself on Adjust approach based on feedback and results

Ten questions, with the answer and explanation after each one.