Analysing prosody requires the correct identification of prominence peaks. This paper examines the results of an experiment where 7 linguists submitted to a prominence identification task largely failed to agree. We show that prominence detection is proportionate to F0 variation, but not to length, that there exists considerable variation between judges, the best of whom barely attains a 50 % score of correct answers. We conclude that coding prosody in a large corpus will require the use of dedicated software to supplement the work done by individual coders.