Lesson 2 · Hands-on Masterclass: Setting Ground Rules to Delivering Your First Data Table Task

This lesson is distilled from the viral practical walkthrough by former unicorn architect and AI tool evaluator @0xlangeai. It takes you step-by-step from signing in, setting cooperation protocols, 4-element prompt crafting, auditing sources against hallucinations, and saving v2 iterations, to mastering advanced capabilities and security perimeters.

🎯 What You Will Master in This Lesson

• Clarify Agent Delivery Boundaries: Avoid vague feature lists; master the tempo of having Muse complete your first tangible task within your visible perimeter;
• Set Up Protocols Right After Login: Master the high-value cooperation protocol prompt and thoroughly inspect permissions, approvals, and existing integrations in Settings;
• Master the 4-Element Task Formula: Adopt the high-hit-rate dispatch structure: Task + Source Scope + Quality Standard + Delivery Format;
• Inspect Outcomes & Verify Sources: Open file deliverables, inspect headers, and cross-check source links to prevent agent hallucinations;
• Iterate and Save as v2 in the Same Chat: Keep conversational context without opening fresh chats to add columns, verify claims, and preserve versions;
• Master Advanced Capabilities & Safety Guardrails: Understand the organic evolution of Memory and Soul, long-term Goals, and the read-only security boundary for local Mac files.

1. Operating Paradigm: From Chat Q&A to Task Closure

Muse is Meta's personal AI Agent capable of autonomously browsing web pages and generating structured files. Long-running tasks continue running in the cloud even after closing the app.

For beginners, the biggest pitfall is immediately delegating a grand and ambiguous goal (e.g., "Analyze the global macroeconomy and give me an investment strategy"). The safest and most effective way to start is to have it organize a few specific URLs into a structured reference table: assign the task, open the file, audit sources, and iterate a new version.

Completing this first loop gives you genuine command over Muse. Connectors, Memory, and Goals can be unlocked gradually as needed.

2. Where to Begin: Login & Scope of Availability

1. Entrances and Device Options

You can begin via any of the following platforms (we use the Muse Mac Desktop Client for demonstration):

• Web Browser: Visit muse.ai and follow prompts to log in;
• Desktop Client: Download and run the native Mac client;
• Mobile App: Download the official Muse app on iOS or Android.

2. Three Facts to Keep in Mind

1. Regional Rollout: Initially launched in the US (iOS, Android, Web), with gradual rollouts in other regions. Account eligibility is based on the registration prompt;
2. Free vs. Subscription: Free tiers and paid subscriptions are available. Run your first real-world task with the free allocation before considering subscriptions;
3. Troubleshooting Blockers: If you cannot enter, check the exact prompt: region waitlist, age verification requirements, or network connectivity.

3. After Login: First Establish How to Cooperate

Upon completing initial naming and setup, never assign complex work right away. Send this【Cooperation Protocol】in your very first message:

I am a user in timezone Asia/Shanghai (UTC+8).
Always provide the conclusion first, followed by necessary explanations.
Explicitly flag any unverified information and attach verifiable sources.

For any external actions—sending messages, submitting forms, placing orders,
making payments, or altering account settings—
always present the exact details and wait for my explicit confirmation first.

Audit Settings & Permission Guardrails

Immediately open Settings to inspect permissions, approvals, and existing connections. Verify:

1. Connected Accounts: Confirm whether connected Google or Meta accounts are primary work or isolated test accounts;
2. Access Scope: Check whether permissions are full access or read-only on demand;
3. Execution Permissions: For modifying calendars or sending emails, ensure "Always ask for confirmation" is strictly enabled.

🚨 Critical Rule: Chat prompts cannot replace system permission toggles! When an approval card pops up, review the specific action before clicking 【Allow】.

4. Practice Task: Organizing Web Links into a Data Table

For your first task, pick a topic you already know well. Only when you are familiar with the topic can you instantly spot whether the agent is hallucinating or misattributing facts.

Here we use Meta's official launch announcement and design specification to organize a reference sheet without needing email or cloud drive connections.

Send the following structured prompt directly to Muse:

Please read the following two links:
1. Official Announcement: https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
2. Design Spec: https://introducing.muse.ai/

Organize a "Muse Core Features Reference Table" with the following requirements:
1. Present in a table containing: Feature Module, Official Highlights, Use Cases, Known Limitations (if any).
2. Base strictly on the text provided; mark any unmentioned points as "Not specified"—no assumptions or extrapolation allowed.
3. Append corresponding source paragraphs or links beneath the table.
4. Upon completion, generate a downloadable file (Markdown or Spreadsheet) and provide the file link.
If file generation is unavailable, provide a complete copyable table and explain limitations.

The 4-Element Task Formula

This task succeeds consistently because it explicitly defines four dimensions: $$ ext{Effective Dispatch} = ext{Core Task} + ext{Source Scope} + ext{Quality Standard} + ext{Delivery Format}$$

5. Upon Receiving Results: Open First, Then Audit

When Muse signals "Completed", never just glance at the chat text. Execute this two-step verification:

1. Open the File Deliverable

Ensure content renders completely, table headers are intact, and links are clickable. If Muse only provided raw text, reply:

I do not see the file link. Please confirm if the file is generated, and provide the exact filename and location.

2. Spot-Check Conclusions Against Sources

Pick 2 or 3 claims to audit:

• Is the claim "continues working after closing the app" backed by official text?
• Are Goals described accurately without exaggerating automation into a guaranteed outcome?
• Are unmentioned pricing tiers honestly marked as "Not specified"?

6. Iterating in the Same Conversation: Save as v2

The first draft is rarely final. Never open a new conversation! Request revisions directly in the current chat:

Please revise based on the previous table. Keep the original file and save the revision as "Muse-Guide-Table-v2".

1. Add a "Publication Date" column; if missing, write "Not specified".
2. Review descriptions against sources; do not frame capabilities as guaranteed results.
3. Move any claims lacking direct source support into an "Unverified Items" section.

Provide the new file link and briefly list what was modified.

Open the new file to verify that new columns appeared and issues were fixed. This completes the full loop of Instruction -> Verification -> Audit -> Iteration.

7. Advanced Capabilities: Add When Needed

1. Memory and Soul: Organic Growth

• Memory: Stores personal preferences, frequent locations, and collaboration habits;
• Soul: Defines communication style and tone.

Do not manually edit these files initially; guide Muse naturally (e.g., *"Provide brief conclusions first, then elaborate if asked"*).

2. Long-Term Goals: Multi-Day Project Tracking

When tasks take several days, use a Goal prompt to define milestones and deliverable requirements.

3. Connectors & Mac Local File Safety

Always adhere to the principle of least privilege:

List filenames in [Full Directory Path].
Do not read unrelated directories. Do not modify, delete, or move files or install software.
Stop after listing and wait for my review.

📌 Summary Takeaway

Managing an Agent is like mentoring an intern: establish protocols first, set clear boundaries, verify sources, and preserve versions across iterations!