This paper investigates how a core component of human interaction – action projection – is realized in conversations with systems like ChatGPT, which are not able to recognize nor anticipate actions in a sequential sense. We ask what this means for users who engage in longer, multi-step projects built through incremental prompt-response pairs. We draw on Conversation Analysis to examine how early adopters of ChatGPT sequentially organise such interactions, using anonymised naturalistic interactions from the WildChat dataset. Our analysis reveals how action projection in LLM interactions is beset with interactional trouble. In turn, we document a set of practices we term ‘pre-emptive prompting’ that users develop to work around the trouble. In doing so, users employ a range of conversational methods including pre-sequences, prompt categorization, withholding executables, spelling out relevant next actions, eliciting displays of “understanding”, and creating slots for redirection in their attempts to complete their projects. Our work provides empirical evidence of how users interactionally accomplish the configuration work (Alcaras and Ricci 2025) needed to work around the discretization and cluttering that LLMs-in-use produce.Preprint, likely to be revised. Comments are welcome!