What it means In the captured material, last-mile filtering is Linus Lee's habit of making the final selection yourself when working with a language model. Rather than asking the model to pick the single best answer, he runs the same prompt several times, notes which options appear across runs, and then personally filters that shortlist down to the answer he actually wants [^3e5745ab-279e-5154-9c22-9015dba70a32@2714] [^3e5745ab-279e-5154-9c22-9015dba70a32@2855]. Why he filters the last mile himself The reasoning he gives is simple: the model doesn't really know you. It can narrow a broad question down to a handful of strong candidates, but the final judgment call — which answer fits your taste, context, and goals — is yours to make. Handing that last step back to the model forfeits agency you don't need to give up. How it fits his prompting habits Last-mile filtering sits alongside the other habits he demonstrates for conversational prompting: start broad with rich context, sample the distribution by re-running the same prompt, and rewrite freely — asking for a redo is iteration, not impoliteness. Filtering the last mile yourself is the closing step of that loop: let the model generate options, but keep the choosing [^3e5745ab-279e-5154-9c22-9015dba70a32@2714]. - Last-mile filtering = doing the final pick yourself instead of asking the model for one best answer [^3e5745ab-279e-5154-9c22-9015dba70a32@2855].- It pairs with sampling the distribution: run the same prompt several times and select from the overlap across runs [^3e5745ab-279e-5154-9c22-9015dba70a32@2714].- The rationale: the model doesn't know you, so the last judgment call should stay with you.