context-engineering

Context engineering reduces hallucinations by selectively loading only the required spec sections, relevant source files, and latest error output into the agent session. This structured hierarchy prevents convention drift and stops the agent from guessing missing requirements.

Context engineering is the practice of structuring and curating the information an AI agent receives during a development session. It establishes a context hierarchy across rules, specs, source files, and test feedback to improve output quality and manage conversation history.

Set up the right context by loading your persistent project rules file, the relevant spec section, the specific source files to change, and only the latest failing test or error output. This prioritization ensures the agent focuses strictly on the current task.

Agents drift from codebase conventions when context grows stale or too large during extended sessions. Context engineering addresses this by recommending session restarts or summarizing managed conversation history to maintain strict adherence to project rules.

Handle missing requirements by establishing trusted versus untrusted context verification expectations. Use explicit confusion handling patterns to surface conflicts and ask for missing requirements instead of allowing the agent to guess or hallucinate solutions.

Restart or summarize an AI coding agent session when the context grows stale, becomes too large, or output quality degrades. This session management step clears managed conversation history and restores the correct context hierarchy for ongoing tasks.