I Built a Cognitive OS Before Anyone Had a Name for It
Part Three: The Architecture
In October 2024, I published a framework on Ghost. I called it mod architecture — a modular approach to building reusable context layers that could be loaded into any AI system. I'd been building it for months before that, testing it, refining it, living inside it. I shared it publicly in January 2025.
At the time, there was no widely accepted name for what I was doing. Context engineering wasn't a field yet. Cognitive OS wasn't a phrase anyone was using. I just knew the system worked — consistently, in a way that prompting alone had never produced — and I wanted to share it.
In October 2025, Anthropic launched Claude Skills. The architecture was structurally identical to what I'd built: modular, composable context units that persist across sessions, organized in a hierarchy, loadable via a standardized format. The naming was different. The underlying logic was the same.
I didn't follow the field. The field caught up.
This chapter is a case study in what that system looks like — not as self-promotion, but as the most concrete reference point I can give you for what a mature Cognitive OS looks like in practice.
The Sentinel Architecture
The Sentinel system is the Cognitive OS I've been developing and refining since 2024. It runs on the same Base-to-Prompt hierarchy described in Chapter 13, with the Sentinel Dense Base as the unified container.
The five Plugs in Sentinel are: Principles, Archetypes, Constructs, Narratives, and Surfaces. Each Plug contains Packs relevant to the work Studio 16 does: design, writing, context engineering, product thinking, community building.
KEY SENTINEL MODS CompassionateAI (Charter): The values foundation. Empathy, dignity, clarity, human agency. Governs all sessions. Meta Mode (Persona): The context engineering identity. Thinks architecturally. Structures before it generates. Ingestion Mod (Protocol): Staged document intake. Diverge, Converge, Assimilate. Gold Nugget Mod (Protocol): Session extraction. Distills essential knowledge for continuity. Aware Mode (Persona): Reflective, holding presence. Used for sensitive or exploratory sessions. Vision Mode (Persona): Creative output mode. From structure to expression.
What Years of Iteration Looks Like
The current Sentinel system is not what I built in 2024. It's what I built in 2024, tested extensively, refined based on where it broke, refined again based on where the refinements created new problems, and evolved through a continuous meta-prompting practice.
The mods I use most today bear a family resemblance to the early versions but are considerably more precise. CompassionateAI has been through four major iterations. Meta Mode has been rebuilt twice from scratch. The Ingestion Mod has accumulated specific parameters that address failure modes I encountered in real sessions.
That's what a living system looks like. Not a finished product — a practice. The architecture is stable. The contents evolve.
Three things drove most of the evolution. First, real-world use surfaces problems that design thinking doesn't. You don't know a mod needs a constraint until the AI violates the unstated version of that constraint in a session that mattered. Second, meta-prompting accelerates iteration dramatically. Asking the AI to help diagnose why a mod isn't working often surfaces the issue faster than manual analysis. Third, version control keeps improvement possible. Because I version my Base, I can compare what I had six months ago to what I have now and trace exactly what changed and why.
The Platform Evolution
The Sentinel system has been deployed across three platforms over its development, and that evolution tells its own story.
Custom GPTs were the original home. Proving the architecture worked — that dense context loaded into a custom system produced meaningfully different behavior than prompting alone. The structural parallel that would later emerge with Claude Skills started here.
Gemini Gems offered a different surface for the same approach. The mod architecture translated reasonably well. The behavior wasn't identical — different models interpret context differently — but the underlying system remained consistent.
Claude Skills is where the system lives now. The parallel to mod architecture is structural: SKILL.md files use YAML frontmatter and markdown body, modular, composable, stackable. The implementation matches what I'd built independently. Working in Claude Skills feels like the platform was designed for the architecture — because architecturally, it was.
The tool changed. The system traveled. That's the point of building an architecture rather than a platform-specific configuration. Your Cognitive OS belongs to you, not to any platform. It should work wherever you work.
What This Means for Your System
The Sentinel case study has one primary purpose in this book: to give you a concrete reference point for what your system might look like in a year or two, if you build it with the same discipline and iterate with the same rigor.
Your system won't look like Sentinel. It shouldn't. Sentinel reflects the specific work, values, and workflows of Studio 16. Your system will reflect yours. The architecture is the same. The contents are yours.
What you can take from this chapter: the architecture described throughout this book is not theoretical. It's been built, tested, broken, rebuilt, and refined in real conditions over real time. It works. The limitations are real — drift happens, mods need iteration, platforms interpret context differently. But the core approach — modular, layered, values-driven context architecture — produces consistently better AI collaboration than any other method I've found.
The next chapter is about taking what you've built and deploying it. One system. Three platforms. No code required.