Video summary
How to Build Your GPT-6 Astra OS
Main summary
Key takeaways
Main Ideas / Lessons
- Organize your AI “operating system” with a clean folder/workspace structure so you can run tasks in isolated areas (e.g., YouTube vs. newsletter vs. client work) without loading irrelevant context.
- Use a single top-level
agents.mdas a “router” that controls overall behavior and routes the model into the right departments/rooms, rather than creating many separate agent instruction files. - Create layered context that scales:
- Company-wide context (shared identity, priorities, rules)
- Department context (e.g., Brand, Agency, Ventures)
- Room context (specific ongoing responsibilities like YouTube, Newsletter, Support, Delivery, or specific client work)
- Decide between “rooms” and
context.mdfiles based on what changes and how much responsibility differs:- If something is an ongoing responsibility with meaningfully different context, give it its own room (and likely its own folder and context files).
- If something is reusable facts/instructions that support an existing room, use a context file.
- Maintain accuracy over time by keeping contexts current and performing audits (daily/weekly/monthly depending on your needs) to prevent “context bleed” and misrouted workflows.
Recommended Methodology
1) Top-Level Structure (the “Company Folder”)
- Create a single main folder for your AI OS (e.g., your company / LLC / operator workspace).
- Inside the top level, include:
agents.md- Acts as system instructions and a router for the whole operation.
- Should set the model’s working role, decision/approval expectations, and routing behavior.
- Shared company context files (stored in a dedicated context area)
- Identity & priorities that apply across the entire business.
- Shared rules that should be consistently enforced.
2) Break Down into Departments
- Under the main folder, create subfolders for major business functions/departments, such as:
- Marketing / Brand work
- Agency / client operations
- Ventures / separate projects/publications
3) Break Departments into “Rooms” (Specific Responsibilities)
- Within each department folder, create “rooms” representing distinct ongoing responsibilities or workflows.
- Example rooms mentioned:
- Brand department: YouTube room, Substack room, LinkedIn room, byline/guest-writer room
- Agency department: client operations room, deployed SaaS/tool repo location
- Ventures: separate publication projects
4) Add Context Files at Each Layer
- Company-wide context files
- General who-you-are/what-you-do/rules/audiences.
- Department context files
- Explanation of that department’s purpose and responsibilities.
- Room context files
- Room description (e.g., a
room.mdexplaining what the room is) - Room-specific workflows and constraints
- Additional context such as audience/ICP specifics per channel or publication
- Room description (e.g., a
5) Implement Routing Behavior via agents.md and Boot Sequence
- In
agents.md, instruct the model to:- Start in an “executive partner” mode (example role: CEO as decision maker; AI surfaces facts/evidence/opinions/recommendations).
- Stay at the root of the company folder and route into branch rooms.
- Do not spawn separate operating sessions purely because a task belongs to a particular department folder.
- Add a boot sequence in
agents.mdthat at the start of each session:- Forces Code(s) to read the key company-wide context files (e.g., a user/profile file, business overview, and ICP/audience profiles).
6) Work Example (How a Session Routes)
- When starting a fresh session in Codeex with GPT-6 Astra:
- The system starts at the company root
- Loads
agents.md+ root context - Detects the task (e.g., “create a YouTube video”)
- Routes into the correct department (Brand) and room (YouTube) context
- Executes using only the relevant loaded context
7) Room vs. Context File Decision Rule
- Use a room when:
- The responsibility is ongoing
- The context/rules differ meaningfully from other responsibilities
- Examples: channels (YouTube vs. Substack), clients (each client likely gets its own room), products/ventures
- Use a context file when:
- The information is reusable facts/instructions
- It supports the existing room’s mission/workflows
- Examples: a specific workflow inside a department (e.g., a cross-channel repurposing contract for YouTube → Substack)
8) Keep Context Organized and Prevent “Context Bleed”
- Ensure clients/channels/products have separate context boundaries (rooms) to avoid mixing instructions and audience/rules.
- Run periodic audits (daily/weekly/monthly) so workflows and context don’t drift into the wrong places.
Concrete Example of the Creator’s Setup
- Main folder: one company-wide workspace (named like “Alex McFarland” in the example).
- Departments inside the company:
- Brand work (audience + educational content)
- Agency (client work + developing a SaaS tool/repo)
- Ventures (building separate publications/projects)
- Rooms under Brand:
- YouTube, Substack, LinkedIn, and byline/guest-writing
- Creator Studio:
- A locally deployed dashboard in Codeex for tracking competitors/creators and building content packages (mentioned as something built alongside the brand work).
Speakers / Sources Featured
- Speaker: The video narrator / author speaking throughout (no name explicitly confirmed in the transcript, but the example “
Alex McFarland” appears to be the speaker’s company/operator name used for the folder template).