Form preparation in word production, the benefit of exploiting a useful common sound (such as the first phoneme) of iteratively spoken small groups of words, is notoriously fastidious, exhibiting a seemingly categorical, all-or-none character and a corresponding susceptibility to "killers" of preparation. In particular, the presence of a single exception item in a group of otherwise phonologically consistent words has been found to eliminate the benefit of knowing a majority characteristic. This has been interpreted to mean that form preparation amounts to partial production and thus provides a window on fundamental processes of phonological word encoding (e.g., Levelt, Roelofs, & Meyer, 1999). However, preparation of only fully distributed properties appears to be nonoptimal and is difficult to reconcile with the sensitivity of cognitive responses to probabilities in other domains. We show here that the all-or-none characteristic of form preparation is specific to task format. Preparation for sets that included an exception item occurred in ecologically valid production tasks, picture naming (Experiment 1), and word naming (Experiment 2). Preparation failed only in the commonly used, but indirect and resource-intensive, associative cuing task (Experiment 3). We outline an account of form preparation in which anticipation of word-initial phonological fragments uses a limited-capacity, sustained attentional capability that points to rather than enacts possibilities for imminent speech.