Why this phase exists
The first two phases produce two extraordinarily valuable artifacts: a validated Problem Statement and a Solution Brief with organizational consensus. Together, they contain the complete human understanding of the problem and the solution. But there is a chasm between what a human team understands and what an AI can process.
That chasm is the context. And Context Engineering is the discipline of crossing it.
When a team asks an AI to build something, the AI knows nothing about the user interviews. It hasn't lived through the synthesis sessions. It doesn't understand the organizational politics that determined a constraint. Everything the AI has is what we give it. And the quality of what it produces is limited — absolutely — by the quality of what it receives.
The operating system metaphor
This phase translates human thinking into a project operating system composed of three inseparable elements:
| Element | What it is | Without it... |
|---|---|---|
| Agents | Orchestrators and executors of context. Specialized AI profiles with a defined role, concrete capabilities, and clear limits. They're not "AI" in the abstract — they are designed roles. | AI is a genius without a role. It can do anything but doesn't know what's expected of it. |
| Rules | General norms that apply to the entire project without exception. Conventions, constraints, cross-cutting decisions. The constitution that no Agent can contradict when executing any Skill. | Each Agent makes locally correct decisions in isolation that are incoherent with each other. |
| Skills | Concrete, specific tasks that execute specific context. Each Skill is a self-contained work unit with all the context an Agent needs to complete it without ambiguity. | Agents have identity and Rules but don't know what to do. They have a role but no tasks. |
Agents without Rules produce incoherent outputs. Rules without Agents are dead documentation. Agents with Rules but without Skills know who they are and what's permitted, but have no concrete work. Skills without Rules produce tasks that don't respect standards. Skills without Agents are instructions without an executor. The three elements are interdependent. Context Engineering is the art of designing all three to function as a system.
This phase is the hinge
Phase 3 is the hinge between pure human thinking (Phases 1 and 2) and AI execution (Phase 4). Everything done poorly here is exponentially amplified when the context reaches the AI. An ambiguous solution produces ambiguous context. Ambiguous context produces output that looks correct but doesn't solve what it was supposed to solve.
Context Engineering is design, not writing. Writing context for AI is not summarizing what you know. It's designing a system of Agents, Rules, and Skills where each piece has a purpose, a calibrated level of precision, and a reason for existing. Poor context produces poor results regardless of which model you use.