Claude Code
Source Leak

Claude Code has a 1M-token context window per session. As your session fills up with messages, tool results, and file contents, it approaches that limit and compacts — summarizing and clearing older context. By default this happens at 95% full. The problem: quality starts degrading well before that, because Claude is trying to reason over a massively loaded context. Setting the threshold to 70% means compaction triggers earlier and more often. Claude gets a cleaner working context. On any session longer than 30 minutes, this is noticeable: more coherent plans, fewer "I lost track of what we were doing" moments. This one change is worth more than any plugin you'll install.

Every time you open Claude Code normally, it loads your entire setup: CLAUDE.md files, memory files, MCP servers, plugins, all of it. That's what takes 3-4 seconds on launch. The --bare flag skips all of that and starts in under 400ms — raw Claude Code with nothing attached. Use it when you just have a quick question that doesn't need your full project context: "What does this error mean?", "Write me a regex for this pattern", "Fix this typo". Keep your full session for real project work where you actually need your memory and tools loaded.

Type /insights in any active session and Claude reads your past 30 days of session history, finds your patterns, and generates an HTML report with specific recommendations. It finds what you spend time on, which tools you underuse, and where you get stuck repeatedly. The person who first documented this found they were spending 40% of every session re-explaining their project setup to Claude from scratch. That's exactly what a CLAUDE.md file should handle — but they didn't have one configured. /insights surfaced that in minutes. One week of fixes dropped their context-setting time to under 10%.

Type /dream and Claude runs a 4-phase process across everything it's learned in your recent sessions: it orients to your current project state, collects learnings and corrections from recent sessions, consolidates them into structured memory files, then prunes and indexes to keep things clean. It runs as a read-only subagent — it can read your project files but can't modify anything during the process. Memory is stored in four types: your role and preferences, corrections you've given Claude, project goals and context, and pointers to external resources. Run it every few days to keep your sessions sharp.

Set this environment variable and one Claude instance becomes a coordinator that can spawn worker Claudes, each running in an isolated git worktree with its own context window. Workers communicate via messages and share a task board with dependency tracking. Use --capacity 3 to limit parallel workers. Start with 2-3 agents on a specific task — 5+ agents on a small codebase creates more coordination overhead than it's worth. Best suited for large tasks with clearly parallelizable subtasks: separate research, implementation, and verification running at the same time.