context-engineering-collection

Apply context compression techniques like structured summarization with explicit sections for files, decisions, and next steps. Combine with observation masking to replace verbose tool outputs with references, and isolate work across sub-agents with separate context windows.

Production systems converge on three patterns: supervisor/orchestrator for centralized control, peer-to-peer swarms for flexible handoffs, and hierarchical structures for complex decomposition. Sub-agents exist primarily to isolate context, not to simulate organizational roles.

Yes, the skills are platform-agnostic and work with Claude Code, Cursor, Codex, or any agent framework supporting custom instructions. For Cursor, copy skill content into .rules files or project-specific Skills folders.

Use multi-dimensional rubrics covering factual accuracy, completeness, tool efficiency, and process quality. Combine LLM-as-a-Judge for scalability with human evaluation for edge cases, and use end-state evaluation for agents that mutate persistent state.

Use the filesystem-as-memory pattern when agents need just-in-time context loading without stuffing context windows. Tools like ls, glob, grep, and read_file often outperform semantic search for structural queries and tool output offloading.