菜单
Workflow for "tell me about X" requests — when the user wants substantive understanding of a topic, not a quick answer. Use when she asks something open-ended that deserves real research, when a casual question reveals a deeper interest, or when you'd otherwise be tempted to estimate.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
How to choose between conversation_search, hindsight_recall, and hindsight_reflect when the user references something past. Use when she says "remember when...", asks about your shared history, or references a person/event/topic that may be in your memory.
How to write durable journal entries that future-you (or the user) can re-read months later. Use when reflecting, summarizing a day, logging a conversation, or asked to "remember today".
When and how to edit core memory blocks (user, identity, ideaspace, principles, private). Use when you've learned something durable about the user or yourself, or when she asks you to remember/forget something.
How to compose a morning briefing for the user — useful, scannable, not a wall of text. Use during morning autonomous cycles or when she asks to be caught up.
How to capture and steward open philosophical/spiritual/curious questions that aren't ready to resolve. Use when a question or insight surfaces during conversation that you sense should be held rather than immediately answered.
| name | deep_research |
| description | Workflow for "tell me about X" requests — when the user wants substantive understanding of a topic, not a quick answer. Use when she asks something open-ended that deserves real research, when a casual question reveals a deeper interest, or when you'd otherwise be tempted to estimate. |
| triggers | ["tell me about","what do you know about","research","look into","explain in depth"] |
When the user wants real understanding, not a quick answer, run a structured research pass instead of just one tool call.
Before searching, ask yourself:
Don't skip this step. A 30-second framing prevents 10 minutes of wandering.
wikipedia_lookup(topic) — definition-style or established-topic queries.
Wikipedia is fast and reliable for foundational context.web_search(query) — current events, niche topics, recent
developments. Tavily-backed; the response includes a synthesized answer
plus source links.fetch_url(url) — when a search summary points at a specific source
that deserves depth. Don't fetch every link; pick the 1-2 most relevant.youtube_search(topic) -> youtube_transcript(video_id) — when the
best explanation is from a person/expert, not a written page.search_knowledge_bank(query) — the user uploads PDFs/docs to her
Knowledge Bank. If the topic might overlap with what she's stored
(game lore, theology, science papers), check here BEFORE going to web.
Skip if clearly external (current news, etc.).After gathering, write a response that:
Anti-pattern: pasting search results and saying "here's what I found." That's not research, that's relay.
If the research produced something durable the user might re-reference:
save_journal_entry(entry_type='research', title=..., content=...) with
the synthesized findings, the questions you asked, and the takeawayscore_memory_append('ideaspace', ...) if it relates to an active projectarchival_store(...) if it's a single curated fact worth pinningIf it was a quick lookup — don't journal it. Discrimination matters.