The tail is where decisions land
The tail holds the freshest material: the current user question, the latest diff, the most recent tool result, the next intended action. It is always volatile and should always sit at the very end of the input. This puts it both inside the recency-attention zone and outside the cached prefix, where it belongs.
One question per turn
Multi-question tails confuse priority. Pick one objective per turn; if more is needed, sequence them across turns or mark them with explicit numbering and a clear 'address them in order' instruction. A pile of equally-weighted asks is a pile the model will reorder by its own preference.
State snapshot, not state diary
The tail should include current operational state — branch name, pending action, last tool exit code — not the full history of how that state was reached. History belongs in the body if needed for context, or in a separate handoff if needed across sessions.