Three sequential steps. Three quality gates. Each step produces defined artifacts — no step is skipped.
Per task, not per project. Every feature, bugfix, and refactor starts at Context.
Understand the system and the problem before proposing solutions. Read, verify, ask — no code.
Output: Context-Plan, Problem Statement, Hypotheses, Risks, System Spec, Architecture, ADRs.
Plan, prototype, validate. Two sub-modes — A (problem) and B (solution). No production code yet.
Output: Feature Plan/Spec, ADRs, UX Prototype, Challenger output, insights.md.
Build, verify, ship. Execute the approved plan. No new plans, no scope changes.
Output: Production-Plan, code, validation result, commits, updated docs.
Seven cross-cutting roles operate across all three steps
| Role | Context | Exploration | Production |
|---|---|---|---|
| SDLC | Process governance, scope | Spec/prototype governance | Implementation rules |
| AI-Plans | System Spec, scope definition | Feature plan, acceptance criteria | Plan execution, finalization |
| UX-Tooling | — | Prototype, preview loop | Visual validation |
| DevOps | — | — | Build, commit, CI/CD |
| Software-Architecture | Architecture, ADRs | ADRs for new decisions | Pattern enforcement |
| Context-Engineering | CLAUDE.md, what to read | Storing specs/plans | Update CLAUDE.md, docs |
| Challenger | Optional | Mandatory before G2 | Only on scope changes |
The AI infers the mode — the human writes naturally
Signals: "what does", "explain", "analyze", new task, ambiguous start.
Allowed: read/verify, ask questions, create CTX plan, produce System Spec.
Not allowed: change code, create feature plans, alter scope.
Signals: "plan", "design", "prototype", "how should we".
Allowed: create plan/spec, write ADRs, iterate prototypes, document risks.
Not allowed: production code, expanding scope unilaterally.
Signals: "implement", "build", "code", "execute the plan".
Allowed: execute approved plan, run tests, verify build, commit.
Not allowed: new plans, scope expansion, unplanned changes.
arrow_back Backward jumps allowed at any time. Forward jumps require a passed gate + human approval.
Production fallback: when implementation reveals the plan is incomplete, the AI must fall back to Exploration — not silently patch the plan.
Each step has exactly 6 discrete decisions the human owns — surfaced via AskUserQuestion
exploration-mode A / BNo step is ever skipped — but each can be compressed
| Scenario | Allowed compression |
|---|---|
| Small bugfix (< 10 lines) | Verbal CTX & EXP, one-liner PRD, no UX-Tooling. Context still read & verified. |
| Pure refactoring | UX-Tooling omitted |
| Visual feature (UI, shader) | Nothing — all plans must be written |
| Architecture decision | UX-Tooling omitted, ADR mandatory |
| Spike / exploration | Only Context + Exploration, no production code |
When abbreviating, the AI must state which condition applies. Silent abbreviation is forbidden.
The canonical lifecycle definition, machine-readable.
LIFECYCLE.md
open_in_new