Self-Ask Prompting

Self-ask prompting has the model explicitly pose and answer its own follow-up sub-questions before committing to a final answer. Introduced by Press et al. (2022), it targets multi-hop questions — ones whose answer requires chaining several facts, like "Who was president when the company that makes X was founded?" The model writes "Follow-up: …", answers it, asks the next follow-up, and so on, making the intermediate reasoning explicit and easy to inspect. The paper also showed a "compositionality gap": models often know each sub-fact yet fail to combine them, and self-ask narrows that gap. For SaaS builders, the real value is that each self-asked question is a natural hook for a tool call — you can route the follow-ups to a search API or your own database, turning self-ask into a lightweight agent loop for question answering. Compared with plain chain-of-thought, its structured question/answer format is easier to parse and to plug external retrieval into. It's overkill for single-fact lookups.