Skill UI

Addresses the critical challenge of managing extremely long conversation histories (millions of tokens) in advanced AI agents, such as coding assistants. This technique moves beyond simple token minimization by implementing structured compression strategies (like Anchored Iterative Summarization) to prevent the loss of critical technical details—such as file paths, decisions, and error messages—ensuring the agent maintains functional context across complex, multi-step tasks. It emphasizes optimizing for 'tokens-per-task' rather than just tokens-per-request.

Context Guardian preserves critical context before Claude Code’s automatic compression, collecting verified snapshots, detecting integrity issues, and briefing transitions so heavy-technical sessions keep every essential decision and dependency intact.

Quick-reference card for the Caveman AI persona, offering modes like Lite, Full, Ultra, and Wenyan styles. Includes skills for terse commit messages, PR reviews, and text compression to save tokens. Supports configuration via environment variables or config files to optimize interaction style.

This skill covers strategies for compressing conversation history in AI agent sessions to manage context window limits. It focuses on optimizing tokens-per-task rather than just tokens-per-request. Techniques include anchored iterative summarization, opaque compression, and regenerative summaries. It also addresses artifact trail integrity and evaluation methods like probe-based testing to ensure critical information like file paths and decisions are preserved during compression.

A context preservation skill that creates snapshots and verifies data integrity before automatic compression. Protects critical project information from being lost during Claude Code's context window compression, ensuring zero data loss for complex technical projects with multiple skills and dependencies.

Context optimization extends effective capacity of limited context windows via compression, masking, caching, and partitioning. It aims to better utilize available capacity rather than increasing window size. Techniques include summarizing context, masking verbose outputs, reusing KV-cache, and partitioning work across sub-agents to reduce costs and latency in AI agent systems.