Capstone part 2: improve, compare and submit your prompt toolkit

Step 6 — Improve systematically

Make one or two changes at a time, re-run on the same test inputs and re-score with the same rubric. Keep every version in your log. Most good capstones go through three to five versions.

Techniques you might use (only where they help): clearer task and deliverable, added context, audience and constraints, a structured layout with delimiters, a role, examples (few-shot), a structured output format, decomposition into steps, uncertainty instructions, or a review step.

Step 7 — Compare initial and final results

Create a comparison table: version, change made, score per test input, notes. Then write a short before-and-after analysis:

• What improved, with evidence (scores and quoted output).
• What did not improve, or got worse.
• Remaining limitations and when a person must check or take over.

Be honest. A capstone that shows a modest, well-evidenced improvement and clearly explains its limits is stronger than one claiming perfection.

Common pitfalls

• Moving the goalposts — changing the test inputs or criteria between versions makes the comparison meaningless.
• Cherry-picking — showing only the best run of the final prompt and the worst run of the initial one.
• Unexplained changes — a final prompt that differs in ten ways from the first, with no record of which change helped.
• Unchecked outputs — scoring an output as "accurate" without verifying its facts.
• Over-fitting — a prompt tuned so tightly to your three inputs that it fails on a fourth. Try one new input at the end as a sanity check.

Step 8 — Document your toolkit

Your toolkit should let someone else use your prompts:

• Template(s) with name, version, purpose, variables and output format.
• Usage notes — inputs expected, how to anonymise them, which tool you tested with.
• Review checklist — what a person must check before using an output.
• Known limitations.

What to submit

The final project question asks for all of the following (you may link to a document for long material):

1. Task statement and why it is appropriate.
2. Initial prompt and its outputs (or representative excerpts).
3. Final prompt(s) and their outputs on the same inputs.
4. Comparison table with rubric scores.
5. Explanation of the improvements and why they worked.
6. Toolkit documentation, including limitations and review checklist.

Your submission is assessed against the published final-project rubric with AI assistance; the platform calculates the score, and you can request a human review. Do not include personal or confidential data.

After the course

Keep your prompt log and toolkit up to date as tools change. Re-test important prompts when you switch tools or models — behaviour can change between versions.