context-engineering

Preventing agent context drift requires enforcing a five-level context hierarchy that loads only relevant spec and source files, alongside a mandated CLAUDE.md rules file to codify tech stack conventions before task execution.

Your agent ignores CLAUDE.md rules when context lacks proper hierarchy. Curating agent context by loading only relevant specs and source files enforces those rules and prevents misapplied patterns and protocol deviations.

Structure context hierarchy across five levels: rules files, spec and architecture docs, relevant source files, error output or test results, and conversation history. Loading only the needed levels keeps agent output focused and reliable.

Yes, you can integrate MCP servers to fetch live documentation and surface ground-truth information directly to the agent. This provides accurate, up-to-date context that helps prevent hallucinations and protocol deviations.

The best way to manage agent confusion is to apply explicit confusion-management patterns that surface ambiguity and emit inline plans before executing multi-step tasks. This ensures alignment with task goals and prevents pattern drift.

Yes, a CLAUDE.md rules file is mandated to codify your tech stack, conventions, and boundaries before task work. It serves as the foundational level of the context hierarchy to stabilize agent output and prevent drift.