Capture a Mac app window as agent context. Use when the user asks to snap, inspect, read, capture, summarize, or work from the current Mac window, a visible app, a selected window, or an app screenshot. Strongly prefer for Appshot-like requests, "what is on this window", "use this app as context", and "grab the window text". SKIP for remote websites/files that can be fetched directly, for controlling the UI, clicking, typing, automating actions, or extracting hidden Google Docs/Sheets/Slides/Gmail canvas text beyond what macOS Accessibility exposes.
Adaptive add-on policy and recovery layer that decides WHEN to compact, prune, snapshot, or fork -- replacing fixed-percent auto-compaction across Claude Code, Codex, and MCP-capable hosts. Trigger on auto-compact timing or damage: "when should I compact", "is it safe to compact now or start a fresh session", "auto-compact fires too early/mid-task", "switching to an unrelated task but the window still has space", "context rot", "answers get worse the longer the session runs", "the agent forgot the plan or my decisions after it summarized", "add a layer on top that manages context without changing the agent", raising autoCompactWindow to give the policy room, or installing/tuning a cross-tool compaction policy or PreCompact hook -- even when "compaction" is never said but the problem is context-window pressure or post-summarization memory loss. Do NOT use to summarize a conversation, build RAG, write a summarization prompt (decides WHEN not HOW), or answer max-context-length trivia.
Load when the user or an agent wants to use Grok/xAI locally with their own SuperGrok / X Premium account instead of a paid xAI API key — chat with Grok, generate images or videos with Grok Imagine, or search X/Twitter posts via Grok. Triggers on "用 Grok 聊天/问 Grok", "Grok 生成图片/视频", "grok imagine", "搜一下推特/X 上怎么说", "grok-cli ...", "log in to Grok with my X Premium", "Grok image/video generation", "search tweets with Grok". Uses browser OAuth (no API key needed); image/video generation is paid. Do NOT load for non-Grok LLMs or for the paid xAI API key SDK workflow.
Load when the user wants to view, sync, or analyze multi-agent activity across multiple devices — Claude Code / Codex / Cursor / Maestri / browser-use sessions captured to a shared GitHub-synced pool, OR wants a distilled per-project context document so a fresh agent in a different tool can pick up where another left off. Triggers on "agentlog X" commands, "what did I do today across all my agents", "show me my pool", "sync the pool", "拉一下另一台机器上的 agent 记录", "看看 vps 上跑的", "跨设备 agent log", "跨工具上下文", "switch from Codex to Claude Code", "AGENTS.md / CLAUDE.md", "项目状态文档", "agent context". Do NOT load for single-machine tweet material capture (use `seed`) or one-off project planning.
Use when the user wants a real data analysis on a dataset — finding what drives a metric, testing whether a difference is real, building a predictive model, reducing many variables to a few, or interpreting what numbers mean for the business. Trigger phrases include "analyze this data", "what drives X", "is this significant", "build a model to predict", "find the factors", "do a proper analysis", "what does this data mean". Skip when the user only wants a chart, only wants raw data pulled with no question, asks a closed factual question with no data, or just wants APA-formatted stats output.
Expert cooking assistant for Chinese and Western cuisine. Provides step-by-step recipes grounded in real data from scraped websites (xiachufang, allrecipes, BBC Good Food, etc.), free APIs (TheMealDB, Open Food Facts), and Kaggle datasets (Food.com, Epicurious). Use when user asks to cook a specific dish, wants suggestions based on ingredients they have, asks about technique/substitution/pairing, or wants nutrition info. Recipes are cached as Markdown files in data/recipes/ with YAML frontmatter — never re-fetch what's already there.
Read the current Claude Code session's raw activity log and help the user synthesize it into tweet material. Use when user says /seed, 'save this for twitter', 'this is tweet material', 'capture this', '记一下', '发推', or wants to turn recent work into a tweet.
Simulate feedback from statistically grounded virtual human populations. Use when the user wants to get product or copy feedback from a panel of diverse personas, predict group opinions or voting behavior, run social-experiment sandboxes, or compute a semantic gradient (SGO) for A/B ranking. Also triggers on mentions of Nemotron-Personas, PersonaHub, synthetic population, generative agents, silicon sample, LLM survey simulation, persuadable middle, SGO, social sandbox, 用虚拟人群测试, 模拟民意, 投票预测模拟.