Muse Spark (Meta)

Prompting Muse Spark

How to prompt Meta's Muse Spark 1.3 for agentic coding and long-horizon knowledge work in Muse Code and the Meta Model API.

Which model to choose

Muse Spark 1.3 ships in two reasoning variants. Use xhigh for daily work because it is available now and pairs an Intelligence Index of 61 with fast output. Use max for the hardest reasoning tasks once Meta completes the extra safety testing that currently gates it.

Muse Spark is proprietary and served through Muse Code and the Meta Model API. Meta lists an open-weights release on its roadmap, so do not plan a local deployment around Muse Spark until that ships.

Prompting in Muse Code

Give Muse Spark an open-ended objective and let it use tools to build context across messy sources. It asks clarifying questions when prompts are ambiguous and confirms before consequential actions, so answer those questions directly instead of restating the whole brief.

For long tasks, state whether you want frequent updates or silent background work. When you steer or interrupt a long thread that holds several tasks, name which task the message targets so the model maps it to the right workflow.

Prompting through the Meta Model API

Route agentic and coding work to Muse Spark 1.3 through the Meta Model API. Earlier reasoning modes are available now while max reasoning arrives after further safety testing.

Expect fewer turns than older Muse Spark models. Meta reports about 20 percent fewer tool calls with 25 percent fewer tokens against 1.2, so a prompt written to micromanage each step fights the model’s training. State the deliverable and the constraints, then let it plan.

Current prompting guidance

• State the objective and the deliverable, not the step sequence.
• Answer clarifying questions directly when they come.
• Name consequential actions up front so confirmations stay quick.
• Keep long-form constraints explicit across multi-step tasks because the model preserves them.
• Label which task a message steers when one thread holds several.
• Judge the model on what it can verify, and expect it to say when it hits a hurdle instead of inventing an outcome.

Examples

Agent brief:

Reconcile the attached vendor quotes against the budget spreadsheet in this
directory. Produce a one-page recommendation with cited line items. Confirm
with me before contacting anyone or changing shared files.

Steering message in a busy thread:

This is about the invoicing task, not the migration task. Pause the migration
work and summarize what is done so far.

Prompt upgrade example:

Stale pattern: “Do step 1, report back, wait for approval, then do step 2, report back, then do step 3.”

Better pattern: “Produce the deliverable end to end. Ask me only when a decision is irreversible or the requirements conflict.”