Forge Goal

Help the user find what they actually want an agent to keep working on, then forge it into one goal prompt that ends — one question at a time, in the language they write in, ending in a block they paste into a fresh session on any agent, with or without a /goal command.

Where this sits: beside /forge-prompt, in no chain. That one forges a prompt for a session a person will steer; this one forges a goal for a run nobody steers. Nobody answers questions mid-run, and whoever judges it — a separate model, the agent itself, or the user in the morning — sees only what the run printed.

The Iron Law

FIND THE WANT BEFORE THE CHECK.
EVERY DONE LINE IS SETTLED BY OUTPUT THE RUN PRINTS.

Two halves, two recorded failures. Skip the first and the goal chases what the repo makes easy to check: asked to make brewlog "better", two unskilled runs offered lint, README and path traversal, one recommended lint — the user wanted the settings page usable on a phone. Skip the second and the run never ends, or ends on its own say-so: Claude Code's /goal judge "doesn't run commands or read files independently", and a goal that "renders correctly at 375px" in a repo with no browser is a line nobody can settle.