Encode Intent

Encode intent once, in durable artifacts, so every agent that follows inherits your decisions without being re-briefed. Each bootstrap phase writes a file that the next phase — and every agent session thereafter — reads as structured input instead of rediscovering from scratch.

The four phases

Phase 1: Product Definition
    Creates: PRODUCT-REQUIREMENTS.md
    Unlocks: Product context for all agents
            │
            ▼
Phase 2: Technical Architecture
    Creates: TECHNICAL-ARCHITECTURE.md
    Unlocks: Tech-stack context, coding standards
            │
            ▼
Phase 3: Agent Specialization
    Updates: Agent configs with project context
    Unlocks: Project-specific agent behavior
            │
            ▼
Phase 4: Workflow Activation
    Creates: GitHub board, branch protection
    Unlocks: Full agent workflow

Why this is the differentiator

Generic agents produce generic results. Most AI tooling treats every session as the first one — you re-explain the product, re-explain the stack, re-explain the conventions. Then you re-relearn whatever the agent forgot.

Gemba Flow refuses that loop. Each phase builds context that the next one inherits. By Phase 4, an agent picking up a ticket already knows your product, your stack, your patterns, and your guardrails — because every previous phase wrote them down where the agent will look.

Your intent isn’t a prompt. It’s the project’s DNA. Every agent inherits it. Every session compounds it. Nothing has to be re-explained because nothing was lost.

Why context compounds

• Front-loaded. The expensive thinking (what to build, how, with what constraints) happens once at the start, not on every ticket.
• Durable. Artifacts live in the repo, version-controlled, reviewable. They cannot be lost between sessions.
• Read by everyone. Every agent role reads the relevant subset. The Worker reads architecture; the PM reads the PRD; the Architect reads both.
• Updatable. When the product changes, you change the artifact. Every subsequent agent session reads the new truth.

The anti-pattern

The opposite of encoding intent is the long system prompt. A wall of instructions retyped or copy-pasted into every session, drifting silently as the product evolves, with no version history and no shared source of truth between agents.

If you have ever tried to keep three agents in sync by editing a system prompt three times — you have lived this anti-pattern. Gemba Flow replaces it with files.