Revising systematically, comparing results and keeping prompt versions

Change one thing at a time

If you change the role, add three examples and rewrite the format all at once, you will not know which change helped — or which one broke something else. Treat revisions like small experiments:

1. Hypothesis — "The output is generic because the audience is missing."
2. Change — add one audience sentence.
3. Test — run on the same inputs as before.
4. Compare — score both versions against the same criteria.
5. Decide — keep, discard, or try the next change.

When you are short on time you can make two related changes together, but note that you did.

Text description

A repeating cycle of 6 stages: 1. Hypothesis: name the likely cause; 2. Change one thing; 3. Test on the same inputs; 4. Compare scores on the same criteria; 5. Decide: keep, discard or try next; 6. Record the version and the change. After the last stage the cycle starts again at the first. Stop when results are reliable for the stakes.

Compare fairly

A fair comparison uses:

• the same inputs (keep a small set of test inputs — two typical and one awkward);
• the same criteria (your checklist);
• several runs per version where it matters, because of randomness;
• the same tool and settings where possible (if the model or settings change, note it).

A simple comparison table in your log:

Keep useful prompt versions

Prompts are work products. Keep the ones that work:

• Give each reusable prompt a name, version number and date.
• Record what it is for, the inputs it expects, its known limitations and the tool it was tested with.
• Keep a short change log ("v3: added sentence limit — fixed long outputs").
• Store prompts somewhere you can find and share them — a document, a notes app or a shared library page.

You will turn your best prompts into templates in Module 9, and your final project is a documented toolkit built from these versions.

When to stop iterating

Stop when the prompt meets your criteria on your test inputs reliably enough for the stakes. A brainstorming prompt can be "good enough" quickly; a prompt whose output reaches customers needs more testing and a human review step. Perfection is not the goal — dependable, checked results are.

Stage 2 assessment

Today's completion rule is the Stage 2 assessment, covering Modules 4–6 (roles and goals, techniques, iteration). It unlocks after days 7–11. You can retake it and request a human review of any result.