What is claude-code-goal-draft-policy?

This project is a skill for Claude Code, the AI coding assistant, that helps you write good completion conditions for its /goal command. The /goal command lets Claude keep working across multiple turns on its own until a condition you set is met, with a small fast model checking after each turn whether that condition has been satisfied. The problem this skill solves is that vague completion conditions cause trouble in two directions: either the work never finishes because the wording is too strict or unclear, or Claude wrongly claims it is done and the checking model believes it, because that model only reads the conversation text and never runs any tools itself. This skill turns a rough request into a single ready to paste goal condition built from five parts: a measurable end state, a real command found in the repository that proves the state was reached, the exact evidence in that command's output that counts as success, a guardrail describing what must not be broken along the way, and a stop clause that prevents endless looping if the goal turns out to be unreachable. It can also review and repair an existing goal condition that is not working as intended. Installation is a simple copy of the skill folder into the personal Claude Code skills directory, after which it activates automatically when you ask for autonomous, verifiable work, or can be called directly by name. The author is unusually transparent about testing: every claim in the documentation is labeled as official, as the author's own method, as an inference, or as tested through small experiments, and the repository includes detailed logs from repeated evaluation rounds comparing results with and without the skill. The most recent recorded comparison shows the skill and an unassisted baseline performing close to evenly, so the author is careful to say the skill's benefit is a trend, not a guarantee. This project is not affiliated with or endorsed by Anthropic, and it is released under the MIT license.