context-engineering

To stop Copilot from ignoring project conventions, curate the right context by organizing durable rules files, specs, and source code so the agent receives consistent stack and coding guidance without prompt flooding.

Context engineering for AI agents is the practice of selecting and layering rules files, specs, source files, and conversation history to prevent context starvation, flooding, and stale guidance during code generation.

Set up Copilot workspace rules by establishing durable project rules files that define your stack, commands, and coding conventions, ensuring the agent consistently applies these conventions across new sessions and task switches.

AI agent output degrades during task switching due to stale guidance and context flooding, which can be fixed by repacking task-specific context with only the relevant rules, source files, and errors for the new feature.

The best way to provide project context is assembling the smallest useful set of rules, specs, source files, and errors for the current task, prioritizing persistent conventions over transient conversation history.

Handle ambiguous requirements by surfacing the ambiguity to the agent, checking project precedents in existing specs, and asking for clarification when requirements conflict or are incomplete to maintain accurate output.