Acceptance criteria per work item that can be turned into tests

Goal

Make "done" for a single work item decidable before the work starts, so that the developer, the reviewer and the person who asked for it agree on the same examples.

Prerequisites

A work item with a stated purpose (who needs what and why), access to the person who can answer domain questions, and a place where criteria live next to the item.

Steps

1. Write the rule in one sentence, then the examples. The Gherkin reference describes this shape: an initial context (Given), an event (When) and an expected outcome (Then), and recommends 3–5 steps per example so that it stays readable as a specification.
2. Cover the main case, each boundary and the failure paths. A criterion set for "limit uploads to 10 MB" has at least: exactly 10 MB accepted, 10 MB plus one byte rejected with the specified message, and the rejection leaving no partial file.
3. Use concrete values, not adjectives: "the balance shows £430" rather than "the balance is correct". Gherkin's own examples contrast "there is money in my account" with "my account has a balance of £430".
4. Keep user interaction out of the Given and keep implementation details out of everything: criteria describe behaviour visible at the boundary the user or client sees.
5. Read each criterion and ask: could a tester who has never seen the code decide pass or fail from this text alone? If not, rewrite.
6. Mark what is deliberately out of scope for this item, so that "you forgot X" is answered by the item, not by memory.
7. Before implementation, the developer proposes any missing example; the requester confirms. Disagreements found here are cheaper than in review.
8. Turn the examples into automated tests where the boundary allows it (API, domain logic), and into a manual checklist where it does not (visual layout). Link the test or checklist from the item.

Expected result

Review and acceptance become a comparison against listed examples; scope creep shows up as a new example that has to be negotiated rather than assumed.

Limits and test basis

Examples cannot enumerate everything; non-functional requirements (latency, accessibility) need a measurable target rather than a scenario. Gherkin tooling is optional; the discipline is in the examples, not in the executor. Method synthesised from the cited reference and common practice; no defect-rate effect is claimed.