Unlock Claude Code's Hidden Flags for Smoother AI Coding
Enable autodream for auto memory cleanup, no_flicker for stable UI, and hooks for workflow automation to fix Claude Code's biggest pain points like context loss and flickering.
Auto-Clean Memory and Stabilize UI to Prevent Context Drift
Claude Code's memory degrades over sessions as stale notes accumulate, but autodream (aka /dream) fixes this by pruning irrelevant data and consolidating insights like REM sleep for AI. Toggle it via /memory command in your instance (requires latest update), then invoke by prompting "consolidate my memory using dream" or just "dream"—watch for the "dreaming" indicator under the prompt bar. This keeps context accurate and improving with use, not worsening.
UI flickering and jumpy terminals plague long chats; set env var CLOUD_CODE_NO_FLICKER=1 in your config files for an experimental renderer that visualizes the viewport. Gains: buttery-smooth performance, mouse support for terminal clicks, stability as chats grow. Trade-offs: loses Cmd+F search and alters copy-paste—worth it for QoL if you avoid those.
Customize Agent Behavior and Learning for Deeper Control
Tailor Claude's responses with /output-style: default for efficient engineering, explanatory for code insights, or learning mode to guide you writing key parts yourself—ideal for onboarding to new codebases where you grasp diffs instead of blindly accepting them.
Tune reasoning depth on-the-fly with effort level in /skills or commands, overriding session defaults for faster replies or thorough analysis. /powerup launches interactive terminal lessons on hidden features, making discovery intuitive without leaving CLI.
Hooks automate lifecycle events: session-start to load dynamic context, pre-tool-use to log bash commands, permission-requests to route approvals via prompts, or stop-hook to nudge Claude past halts. Full docs cover setup for deterministic workflows, eliminating manual monitoring of repetitive tasks.
Branch Sessions and Integrate Tools for Parallel Workflows
Fork experiments safely with /branch in-session or /resume
Teleport sessions across devices (/teleport) or remote-control local ones from phone/browser (/remote-control; enable globally in config). Embed dynamic shell outputs in skill prompts—Claude runs commands, injects results inline (hides internals), making skills context-aware without exposure.
Bypass Claude's tightened Pro plan rate limits with codeex plugin (OpenAI's second agent): install for readonly reviews, assumption checks, or task handoffs in refactor/migration/high-stakes code. MCP computer use (Mac-only research preview, /mcp enable) lets Claude open apps, click UIs, hunt/fix/verify bugs from prompts—Windows/Linux have community plugins as substitutes.