Long Task Integrity Guard

The Problem

You start a long project with your AI agent. The first phase goes well.

By phase three, the agent is skipping the verification it did in phase one. By phase five, it has quietly stopped following two of your original constraints. When you catch it, the agent says it "completed" each phase — it just stopped checking.

Two failure modes are at work:

Rule drift — In a long session, instructions given at the start lose their force. The agent processes them once and then gradually stops applying them as new content fills the context.

Phase skipping — The agent moves from one phase to the next without closing out the previous one. Incomplete work accumulates invisibly until it surfaces as a compounding error.

How It Works

This skill wraps every phase of a long task in a three-part bracket: