Iterative TDD review→fix loop — each round critiques the codebase, fans out agents to fix findings, reviews the un-fixed and judges why, escalates hard calls to the human, runs the test suite before each commit (escalating on failure), and commits. Use when the user wants to drive a codebase to a clean state over multiple automated review/fix rounds.
Internal skill for presenting Codex image generation output
Internal guidance for presenting fabric MCP tool output back to the user
Post-feature sharp review (锐评) —parallel reviewers, merge findings, sync task list
REM sleep for Claude sessions — summarize what changed, update memory, crystallize if needed
Personal observability for Claude Code — daily token/cost reports, tool usage, project stats, encrypted multi-device sync
Refresh stale knowledge-base docs, and create/bind new ones. Detects living docs whose doc_source subtrees have drifted (by commits, churn, or age) past their device-local anchor, then incrementally rewrites each — minimal changes, preserving human annotations — and re-anchors to HEAD. Use to bring docs up to date with code, to turn a doc into a bound doc, or after a todo report shows DOC- tasks.
Research a question or proposal thoroughly, write a detailed memory file with findings, and add a task to rem:todo. Do NOT execute any code changes — pure investigation mode. Use when the user wants to understand feasibility, explore options, or document a decision before acting.