/plan, /grill, /goal: the built-in skills workflow
Meta’s one-line pitch for Muse Code’s bundled skills: put a rough idea in, get a grilled, taste-checked feature out. Here’s what each command actually does and when to reach for it.
/plan — nothing changes until you approve
/plan turns a requested task into an approval-gated plan. The agent studies the repository, decomposes the work, and presents its intended approach — and then stops. No code changes until you sign off. For anything beyond a trivial fix, this is the correct opening move: it converts “I hope the agent understood me” into a document you can read, and it catches misunderstandings when they cost seconds instead of an afternoon of unwinding a wrong-headed refactor.
/grill — adversarial review before execution
/grill stress-tests the plan until it holds up. The name is apt: instead of politely accepting the first plausible-looking plan, the agent interrogates its own approach — edge cases, hidden dependencies, failure modes — and revises until the plan survives the grilling. Human engineers do this in design review. Most coding agents skip it entirely and pay for it in retries. If a plan is going to fall apart, you want it to fall apart on paper.
/goal — the finish line, remembered
/goal works toward successful completion of the objective you specify — and keeps the agent on task until what you originally wanted is what gets merged. Long agent sessions have a well-known failure mode: drift. Fifty tool calls in, the agent is dutifully polishing something adjacent to, but not quite, the thing you asked for. Goal conditioning was trained into Muse Spark 1.2 precisely for this, and /goal is the handle you hold it by.
The supporting cast: /model and /effort
Two more commands manage cost rather than workflow. /model switches the backing model mid-session — for instance to muse-spark-1.2 at standard Model API pricing when you need the no-training-data tier (the details are in our tier comparison). /effort dials reasoning depth up or down. Muse Spark thinks before it answers and those thinking tokens bill as output, so matching effort to task difficulty is the single easiest cost optimization available.
A workflow that composes
The skills chain naturally: /plan the feature, /grill the plan, approve it, then let /goal carry execution — with fan-out parallelizing the independent parts and the event log recording every step for review. Each stage gates the next, which is exactly how you’d want an autonomous system to earn incremental trust. The bundled-skills cookbook recipe shows the whole pipeline on a real feature; find it in the cookbook.