Goal Prompt helps write the condition behind /goal, the command that keeps an agent working until a completion condition is met. It drafts an end state that can be true or false, names the files to read first, the exact check command and its expected result, what must not change, and when to stop.
It then closes the cheapest easy-outs, such as deleting code so a search check passes, and formats the result as a single copy-ready line under the 4,000 character limit.
What You Get
- A single-line /goal command in a fenced text block
- A Missing: list for any element that could not be grounded in your request or repository
- Checks paired so the cheapest shortcuts stop working
- Advice to move long briefs into a GOAL.md and a progress log file
Common Use Cases
- Turning a vague task such as a migration into a goal with a test command and a zero-matches check
- Rewriting an existing goal that loops forever
- Setting up a long-running audit goal with a threat model file
- Compressing a long brief to fit the character limit
What Makes It Unique
It focuses on how a goal is judged: in Claude Code a small model re-judges the condition from the transcript alone after each turn, so the skill requires checks to be run with output shown and joins conditions with AND. It never invents missing pieces and lists gaps for you to fill, such as the metric or threshold behind “faster”. A security section collapses audit goals to one terminating criterion and asks for a second validation pass by a fresh agent.
Under The Hood
The skill is prose plus a small Python script that the agent runs through uv in the shell. The script reads a draft from a file or stdin, collapses whitespace, strips a leading /goal, quotes and fences, checks for a stop clause and the length limit, and prints the result. It makes no network calls. The agent may read repository files such as package.json to ground the check command, and may write a draft file.