Anthony comments:

I worked on a project where we got a lot of value by writing the acceptance test for a story before starting to code it. The tests were described in English, and run manually by an independent testing team

Fair point. In (the fabulous) User Stories Applied, Mike Cohn suggests writing acceptance tests on the reverse of your index cards, and I can see how they would help. I’d still maintain that those tests should still be written as close to the beginning of the story as possible, the last responsible moment.

Also, unless regularly run run the likelihood that an acceptance accurately describes the system rapidly decreases once the story is done; manual tests make poor long-term documentation.