context-engineering

To fix AI agent context drift, you need to load hierarchical context models with rules files, architecture docs, and per-task data, enforcing project-specific conventions to keep agent output consistent across sessions.

AI agents hallucinate during task switching when loaded content is too broad. You can prevent this by selectively limiting source access to only task-relevant files and rebuilding context before starting new work.

You can persist project rules across sessions by maintaining rule files like CLAUDE.md that are loaded first in a hierarchical context model, ensuring guidelines are enforced before generating code.

The best way to structure AI context loading is hierarchically: load project-wide rules first, then per-feature specifications, and finally task-specific source files to maintain alignment with architecture conventions.

You should trigger a context refresh when starting a new coding session, when agent output quality declines, or when switching between tasks, allowing the system to rebuild and validate the active context.

No, the system selectively limits loaded content to relevant sources for the current task automatically, applying trust-level controls to avoid hallucinations without requiring manual file selection.