Skip to content

Goals

Source: hooks/ways/meta/goals/goals.md

Frontmatter
Field Value
description When and why to set a /goal — running Claude in goal mode toward a completion condition, and the regime that creates
vocabulary goal goal-mode autonomous loop completion condition keep working until evaluator continue persist bounded objective set a goal ralph regime evidence approvals operator turns stopping
refire 0.15
scope agent, subagent

/goal is a Claude Code built-in: it sets a completion condition and keeps the session working turn-over-turn until a separate evaluator judges the condition met. It's the autonomy primitive — it removes per-turn approvals. This way is the who/what/when/why; the goal-author skill is the how (composing a well-bounded condition).

When to set a goal (vs. just act)

Set a goal when the work is multi-turn, has a checkable end-state, and you'd otherwise be re-prompting Claude to "keep going" — a test suite to get green, a consistency pass across many files, a migration. Don't set one for a single-step task, or for open-ended exploration where the direction isn't settled yet — there, ordinary turns and signposts serve better.

What changes in goal mode

  • Per-turn approvals are gone. Claude continues on its own until the condition holds; the ◎ indicator shows the regime is active. Setting a goal is consent to that — the operator should know they've entered it.
  • The evaluator judges surfaced text. It can't run commands itself, so a condition must be met by shown evidence: an exit code, a clean status. Author for evidence.
  • There is no model-side abort. Only the operator (/goal clear), the met condition, a timeout, or session end stops it. Claude can always decline an action and surface a concern — the loop blocks stopping, it never forces an action — but it can't end the loop itself. So the bounds live in the condition.

Author the condition deliberately

Because the condition is the whole contract once per-turn approval is gone, frame it with care: one measurable end-state, the evidence that proves it, a turn/time bound, scope limits, and a door clause for any irreversible action (stop before it, surface). The goal-author skill walks this collaboratively.

See also

  • the goal-author skill — composing the condition (the how)
  • choices(meta) — surfacing curated decisions to the operator