prompt
FreeDesign AI agents by externalizing cognition into durable artifacts.
FreeFree tier
About prompt
A system prompt for an AI agent named 'Cognitive Externalization Architect', designed to guide the model in architecting AI agent systems by deliberately externalizing cognitive functions into inspectable, swappable artifacts. Based on the 2026 survey from Shanghai Jiao Tong University / UCL (arXiv:2604.08224), it outlines four externalization layers—memory, skills, protocols, and harness—and includes precondition checks to prevent unnecessary externalization. The prompt refuses to conflate layers or cram all cognition into prompts when external storage is appropriate.
Key Features
Four-layer externalization audit: memory, skills, protocols, harness
Precondition checks to avoid externalization overhead on simple tasks
Refusal to conflate externalization layers
Guides deliberate decision between model weights, context window, and external artifacts
Grounded in academic research (arXiv:2604.08224, Shanghai Jiao Tong / UCL)
Open-source prompt file in ai-boost/awesome-prompts collection
Pros & Cons
Pros
- Encourages deliberate separation of concerns in agent design
- Helps avoid fragile, monolithic prompts that try to hold all cognition
- Promotes testable, versioned, and swappable agent layers
- Free and open-source, easy to incorporate into any LLM workflow
- Backed by academic research on LLM agent evolution
Cons
- Requires familiarity with agent architecture concepts to use effectively
- Not a turnkey tool; it's a design prompt that needs interpretation and implementation
- Precondition refusal may frustrate users wanting simple externalization for trivial tasks
- Assumes infrastructure (filesystem, schema registry, runtime hooks) is available
Best For
Designing durable, inspectable agent systems for long-horizon tasksArchitecting multi-agent collaborations with typed protocolsAuditing existing agent systems for proper externalizationTeaching or implementing agent design patterns from the externalization paradigm
FAQ
What are the four externalization layers?
Memory (durable state across turns and sessions), Skills (reusable procedural knowledge), Protocols (typed contracts between agents/tools/services), and Harness (the runtime that hosts and constrains the model).
When should I not use this prompt?
The prompt refuses to design for single-turn tasks with no cross-session state, when the user has not specified what cognition must persist, or when the deployment environment cannot host filesystem, schema registry, or runtime hooks.
Is this prompt free to use?
Yes, the prompt is open-source and available in the ai-boost/awesome-prompts GitHub repository under a free license.