Deep critical analysis of one specific paper — dissect experimental setup, extract numbers, evaluate claims against a hypothesis. Triggers: 'analyze this paper', 'deep dive into', '分析这篇论文', '论文深度分析', 'what does this paper really show', '这篇论文说明了什么', arXiv ID with analysis intent. Not for paper discovery (use scholar-agent), quick summaries, or literature surveys.
Check whether documentation needs updating after code changes. Triggered when user says "sync docs", "文档同步", or "检查文档".
Manage ML experiment lifecycle via structured YAML registry — register, record results, compare runs, track status. Triggers: 'register experiment', 'compare experiments', 'experiment results', 'exp init', 'exp list', '注册实验', '对比实验', '查实验', '哪个实验最好', '记录结果'. Not for training monitoring, checkpoints, or launching jobs.
初始化新项目的 Claude Code 配置。用户说"初始化项目"、"init project"、"配置 Claude Code"时触发。
Pick the next executable task from project-root TODO.md by priority + dependency, mark it in-progress, then start executing. Explicit invocation only — never auto-run. Triggers: 'work on todo', 'next task', 'pick a todo', '领 todo', 'todo worker', '接下来做什么', '领下一个任务'. NOT for: listing TODO contents (just read it), brainstorming what to do, interrupting active coding.
Self-review for paper paragraphs — diagnose v1 issues and give a revision direction (intro, abstract, method, related work). Triggers: 'review this intro', 'check this paragraph', 'paper paragraph self-review', 'help me look at intro / abstract', 'polish this intro', 'paper self-review', '帮我看 intro', 'paper 段落自检', 'review 这段 paper'. Not for: generating a draft from scratch — write v1 yourself first, then iterate.
Audit project context management (CLAUDE.md, rules, knowledge) for progressive disclosure compliance. Checks that all knowledge files are reachable from rules, detects stale references, orphaned files, and CLAUDE.md index leakage. Triggers: 'audit context', 'check context', 'context hygiene', 'context audit', '检查 context', '审计上下文', 'knowledge 覆盖检查'. NOT for: code review, rule content correctness, or knowledge content quality.
Cross-model second opinion from OpenAI Codex on plans or code. Auto-detects plan vs code review, iterates up to 5 rounds. Use for architecture decisions, non-trivial refactors, critical config changes. Triggers: '/codex-review', 'codex review', 'second opinion', 'cross-check', '让 codex 看看', '交叉审查'. Not for trivial fixes or formatting.