General skillAI and agent development
claude-context-window-management
Keep long-running Claude Code sessions within a small context budget
Keep long-running Claude Code sessions within a small context budget
Context Window Management adds operational rules for long Claude Code runs: hard caps on subagent handoffs, a marker-line protocol that accounts for injection and truncation, a launch-turn barrier before merging results, and eight named anti-patterns. Use it before kickoff only when at least two conditions hold: raw tool output is large, several agents reuse the same artifact, or three or more parallel branches need to be merged. If only one condition holds, the project recommends a flat plan instead.
Resource types
Use cases
Platforms
Capabilities
Turn detailed business processes into reusable, tested AI skills.
Define portable task-specific sub-agents in Markdown for several coding assistants.
Audience
Public GitHub facts last synced Jul 10, 2026.
Find, create, run, and improve agent skills from real execution feedback.
Reduce LLM token use by pruning context, caching prompts, and routing models.