---
name: goal-brief
description: Turn a one-line task into a proper goal — finish line, evidence, constraints, stop condition — then start it with /goal. Refuses to start without named evidence.
whenToUse: when the user gives a task that has a finish line but no stated way to verify it, or says "make it a goal"
disableModelInvocation: true
arguments: [task]
---

Place at `.agents/skills/goal-brief/SKILL.md` (project) or `~/.agents/skills/goal-brief/SKILL.md`
(user). Invoke as `/goal-brief <one line>`. The point of this skill is the EVIDENCE
line: a goal without it is checked against the model's opinion instead of a command.

---

The user wants: **$task**

Before starting, produce a goal in exactly this shape. Ask for anything you cannot
fill in from the request, the repository, or `AGENTS.md` — but do not ask obvious
questions, and do not start until every line is filled.

```
FINISH LINE   what must become true, in one sentence
EVIDENCE      the exact command that proves it, and the output you expect
              (prefer a single test file or filter for the loop; name the
              full-suite command for the final check)
CONSTRAINTS   what must not change; what is out of scope
STOP          "stop and report blocked after N turns" — pick N from the size of the task
```

Rules:

- If you cannot name EVIDENCE, propose one from the repository's test setup. If
  there is genuinely no way to verify the work, say so and stop — that is a problem
  to solve before the task, not after.
- Restate the four lines, then start goal mode yourself by sending the objective as
  a single `/goal` line that contains all four — Kimi checks the goal after every
  turn against exactly that text.
- When the goal completes, paste the EVIDENCE command's output verbatim in the
  final message. Do not summarise it. Do not say "tests pass" without the output
  that shows it.