← ClaudeAtlas

go-goal-golisted

Help users write /goal objectives as concise natural-language prose in their own language, with a built-in Reviewer sub-agent check. Triggers on user requests in any language asking to write a goal.
xiehuacheng/skills · ★ 1 · AI & Automation · score 77
Install: claude install-skill xiehuacheng/skills
# go-goal-go A `/goal` plan is **prose in the user's language**, plus one `<reviewer>` block for independent verification. One standard shape, no tier picking, no dense parameter forms. ## Rules - Drafting is read-only until `CreateGoal` is called. All shipping actions need explicit approval. - The Reviewer is a separate sub-agent (not the main loop). On `FAIL` route back to the main loop iteration, decrement remaining budget; do not declare complete. On Reviewer dispatch error (timeout / OOM / refused / empty) surface to the user as a soft pause with the captured partial output, never as a code change. - Default Reviewer role: `verifier`. Default budget: 50 iterations / 30 min wall-clock / $2 / 5-no-progress / 20% reserve for handoff. Budget exhausted terminates with a cost summary; never claim `PASS`. - Default boundaries: project-only paths, write allowed in-scope, deny-all network. - A `<reviewer>` block must be preceded by a one-line **dispatch mandate** naming the sub-agent and forbidding in-place self-judging. See `references/reviewer-template.md` for canonical phrasings. - Verdict parsing: `^(PASS|FAIL:.{1,280})$` on the last non-blank line. Anything else = `REVIEW_INFRA_ERROR`. ## When to use / not Use when the user wants to write a goal or describes a multi-step task that could run unattended. Skip for single-turn questions, exploration with no finish line, or any task where neither proof nor budget can be named. ## Workflow 1. **Fit check.** Queue-shaped + v