Context Discipline is a procedural skill for AI agents that enforces a strict, artifact-gated workflow at every step of execution. Rather than letting an agent silently skip steps, make unverified assumptions, or declare a task complete without evidence, this skill requires agents to inventory available context, declare assumptions explicitly, and produce concrete artifacts as proof before any workflow transition is allowed.
The core mechanism treats each step boundary as a gated edge: an agent cannot advance until it can supply the required evidence. If that evidence is missing — because context is incomplete or an operation has not been verified — the agent stops and requests clarification rather than guessing. This is a fail-closed design: ambiguity causes a halt, not a guess.
For destructive operations such as deletions or security changes, the skill mandates a human-in-the-loop approval gate before proceeding. For multi-step coding workflows, it requires that tests pass and changes are recorded before the agent claims completion. For agent handoffs, it standardizes a “state of the world” summary so the receiving agent has full, auditable context. Unverified actions are automatically labeled as partial or not_run rather than succeeded.
This skill is aimed at developers building complex agentic pipelines where auditability and safety are non-negotiable requirements. It is a structured decision matrix and artifact schema, not a simple prompt — making it more robust than ad-hoc instructions for controlling agent behavior in production workflows.