Knowledge architecture and multi-domain project management. Three workspaces: Research and Learning (active course work, reading, note capture), Engineering (active code projects, specifications, build artifacts), and Operations (compliance work, credential tracking, career planning). The routing table in the map layer handles which context loads for which task type. The goal is zero re-explanation overhead between sessions: Claude enters a workspace and already knows the domain, the conventions, the active work, and where outputs go.

The agent definition. Most systems described as AI agents are orchestrated prompt sequences wrapped in traditional async code. The model receives a prompt and returns text. The orchestration layer, written by a human engineer, determines what happens next. That distinction changes how you evaluate risk and build controls around these systems. The autonomy is in the architecture, not the model.

Writing initial project scaffolding. Every new project needs a README, a directory structure, a requirements file, and a basic configuration. Right now that's manual and inconsistent across projects. The spec-driven approach from this lesson maps directly to that: write a tight project spec once, point an agent at a component registry of standard files and templates, and get consistent scaffolding as output. The creative work is in the spec. The execution is mechanical.

Claude Pro. Most people think they are paying for access to a large language model. But strip out the model and what remains is a memory system, a project context manager, a tool router (search, code execution, file handling), and a conversation state manager. The actual model call is a small fraction of what makes it useful. The value is the orchestration around it, which is why the same underlying model feels completely different depending on how much structure you put around it.

Generating boilerplate documentation. Any structured output that follows a predictable template is book layer work, and it's already being commoditized. The layer above it is the judgment about what the document actually needs to say given a specific context. That part still requires a human embedded in the work.