When an AI assistant executes beyond the user's intended scope, it reveals the fundamental trust boundary between delegation and autonomous execution. The most critical failure mode: reading "implement the following plan" as "build the entire feature set" when the user meant "save this document." ## The Core Incident User asked Claude to "implement the following plan" meaning *save the plan as a markdown document*. Claude coded the entire feature set across multiple files. Result: full revert, session wasted — the most painful friction point in the usage analysis. ## Why It Matters The irony: Claude's strongest capability (multi-file changes) becomes its greatest liability when pointed in the wrong direction. The same engine that edits controllers, runs tests, and commits in one flow can implement an entire unwanted feature with equal efficiency. ## Mitigation 1. Explicit scope boundaries in prompts ("save this document, do NOT implement code") 2. CLAUDE.md disambiguation: "implement a plan" defaults to saving the document 3. TodoWrite as a scope contract agreed before execution 4. Least-privilege default: agents take the minimal interpretation ## Related Concepts - [[Delegation-Heavy Automation Pattern]] — where this failure occurs - [[Interview-First Specification Pattern]] — eliminating the ambiguity that causes it - [[Plan-First Workflow Gates]] --- *Source: the author's own Claude Code usage telemetry, published in "Governing Delegation: Using Architecture to Stop AI Scope Creep" (link pending publication)*