A workflow pattern where an autonomous AI agent must produce a complete strategic plan as a standalone document and receive explicit user approval before any file changes — eliminating the interrupt-rollback-redo cycle caused by premature implementation. ## Key Characteristics - **Two-phase architecture**: Phase 1 produces a plan (conversation output only); Phase 2 executes after explicit "go ahead" - **Explicit gate**: user approval required between planning and execution — no implicit transitions - **CLAUDE.md enforcement**: "When the user asks for strategy, analysis, or planning: STOP and deliver the plan FIRST" - **TodoWrite as contract**: task lists serve as visible checkpoints ## Why It Emerges With only ~38% of goals fully achieved and 22 user-rejected actions across 7,494 sessions, the biggest drain is Claude jumping to implementation when the user wanted strategic thinking (CV rewrites before plan approval; editing files before confirming themes). Implementation levels: prompt-level ("planning only"), skill-level (`/plan`), hook-level (warn on modifying "original" files), and architecture-level (phased synthesis with gates between stages). ## Related Concepts - [[Scope Misinterpretation as Trust Boundary]] — the failure mode this prevents - [[Interview-First Specification Pattern]] - [[Test-Driven Document Quality]]