A successful demo tells you what a skill can do with one clean input. Before giving it real files, make three small fixtures that reveal where you need a human check.
- A clean fixture. Make a synthetic invoice with a date, an invoice number and a total. Write down the expected output before running the skill.
- A missing field. Remove the date. Check that the skill leaves it blank or flags it for review. An invented date should fail the test.
- An ambiguous value. Add two totals with clear labels. Check whether the skill picks the requested field and preserves the original value for review.
Run each fixture in a scratch folder and compare the results with the same expectations each time. Keep the inputs and expected outputs together so you can repeat the check after an update.
Browse the Skillπ AI skill marketplace for skills to evaluate.
