Connection to Phase 4: AI Build Phase
When the Gate Review is complete, the team has a system of Agents with identity and limits, Rules that govern all decisions, and self-contained Skills that define each task. All validated, all traceable.
Construction starts with context, not a briefing
Phase 4 doesn't start with a briefing. It starts with context. And with a high-quality system of Agents, Rules, and Skills, construction tends to be the phase of least cost and least friction. The focus isn't managing the construction — it's protecting the integrity of the tripod during it.
Each decision by a development Agent is evaluated against a single question: "Is this decision faithful to the Rules and the Skill I'm executing?" If it is, it proceeds. If it isn't, it's documented in the Decision Log and escalated to the Context Engineer. Construction doesn't silently modify the Rules. It respects them or updates them explicitly.
The context as a living system
The context doesn't freeze when Phase 4 begins. The three elements continue to evolve:
- Construction reveals ambiguities → Skills are refined.
- A local decision during a Skill that should be global → becomes a new Rule.
- An Agent needs unforeseen capabilities → the Agent definition is updated.
- The market (Phase 5) changes the understanding of the problem → Rules and Skills are updated.
- An active assumption is refuted → Rules and affected Skills are adjusted.
Every change is documented in the Decision Log. Every evolution is traced. The system is alive, but every mutation is deliberate and recorded.
What travels from Phase 3 to Phase 4
| What travels | What it contains | How Phase 4 uses it |
|---|---|---|
| project-context.md (Rules) | Technology stack, conventions, constraints, Agent limits, gap protocol, global decisions. | Every development Agent loads it before executing any Skill. It's the base constraint. |
| Agent definitions | Identity, responsibilities, limits, applicable Rules, gap protocol. | Each Agent knows its role, its boundaries, and what to do when information is missing. |
| PRD | What to build, for whom, use cases, priorities, success criteria. | Guides the sequence and scope of Skills. The QA Agent uses success criteria for validation. |
| Architecture Document | Technical decisions, ADRs, data model, AI strategy, non-functional requirements. | Development Agents apply architecture decisions. The Architect Agent's decisions are Rules. |
| Story Files (Skills) | Self-contained tasks with objective, context, acceptance criteria, dependencies, constraints. | Each Skill is one interaction. The development Agent receives it, applies Rules, produces output. |
| Decision Log | Every change to Rules, Agents, or Skills since their creation. | Traceability. If a question arises about "why is this rule here?", the answer is in the log. |
The complete traceability chain
The full chain is: Problem Statement → Solution Brief → PRD → Rules + Skills. If a Skill can't be traced back to the Problem Statement, it's invention. If a Rule doesn't trace to an architecture decision, it's an assumption. Traceability is not bureaucracy — it's the mechanism that ensures the AI builds what was agreed upon, not what it infers.
AI is only as good as the context it receives. The most powerful model in the world with poor Rules, poorly defined Agents, and incomplete Skills will produce poor results. A competent model with an excellent system of Agents, Rules, and Skills will produce results no human team could match in speed or consistency. The difference was never in the AI. It was always in what we give it.
Context is not documentation. It's the operating system that makes AI think with your head.