The Midterm Goal Pattern decomposes a multi-week deliverable into roughly six binary-outcome mini-goals, each split between AI-owned /goal sprints and explicit human handshakes, so a long objective converges without stalling on actions only the human can take.
The pattern emerges from a structural limit of the 📝Ralph Loop commands shipping in spring 2026 — 📝Claude Code's 📝/goal and 📝Hermes Agent's 📝/goal. Each implementation works reliably for tasks scoped to roughly twenty turns with no human input mid-loop, but any objective worth a multi-week commitment violates both constraints — it spans dozens or hundreds of turns and depends on outputs only the human can produce, such as a recorded video, an approved positioning line, or a credential the agent cannot generate on its own. A single /goal run against such a target either reports premature judge-true success or stalls.
The Midterm Goal Pattern resolves this by inverting the planning layer. Before any AI sprint runs, an orchestrator agent interviews the human to map real context — existing assets, audience size, platform access, hard constraints, hand-offable artifacts — and only then decomposes the midterm objective into mini-goals. Each mini-goal carries a binary outcome the judge can evaluate, an AI portion suitable for one /goal sprint, and a human portion declared up front as a handshake with a specific deliverable the next sprint will consume. The result is a shippable midterm objective on a three-to-four-week cadence with the agent acting on mapped context rather than guessed state, and with the human's role made explicit rather than assumed away.