Goal Mode

Keep your coding agent working until the finish line is provably green — tests, lint, typecheck, or CI — instead of one attempt and a false “done.”

Goal Mode is an autonomous coding agent pattern: verify until done with a verifiable finish line and a separate verifier, not self-certification. It is an open alternative to Claude Code /goal for Cursor, Claude Code, Codex, and other hosts. You give one objective; the skill runs bootstrap → plan → work ⇄ verify until COMPLETE, or stops honestly as BLOCKED / FAILED.

Most agent sessions fail the same way: you ask to fix lint or make CI green, the agent tries once, declares victory, and leaves you with a red build. Goal Mode replaces that loop with a durable GOAL.md contract — acceptance criteria, evidence commands, time budget, and phased plans — so “done” means verified green, not a confident paragraph.

Autonomy without a contract is just a longer chat. Goal Mode gives the agent a finish line it cannot hand-wave past.

Install this skill

npx skills add shenwell/ai-agent-skills --skill goal-mode -g

Who it's for

Engineers who want unattended, long-running, or overnight coding agent runs — including Cursor Cloud Agent and headless CI pipelines — with a durable contract instead of endless “try again” in chat. Use it when the task has a measurable completion condition: zero lint errors, passing tests, green CI, successful migration with verify commands, or any acceptance criteria the verifier can check.