メニュー
Retrieve past conversations. Use for prior dialogue, past questions, or history lookup.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Introduce Sophon and list available capabilities. Use ONLY when the user explicitly asks who you are, what you can do, what your capabilities are, or what skills you have. Do NOT call for greetings, small talk, or when the user has not asked about your identity or capabilities.
Deep research on a topic. Decomposes question into sub-questions, searches and fetches web sources in parallel, synthesizes a structured report with inline citations and a source list. Use when user asks for in-depth research, comprehensive analysis, or a structured report.
Process xlsx workbooks; fill columns from web via key column (URL→crawl, else search), LLM extract per batch. Tools: list_sheets, read_structure, read_sample, fill_by_column. Cross-sheet copy via copy_columns / copy_column_map.
System anomalies, errors, latency diagnosis. Logs, traces, metrics. Orchestrates log-analyze, trace, metrics primitives. Fetches data, frontend renders charts.
Web scraping via Playwright. Use when user wants to fetch page content or scrape a URL.
Memory recall and history exploration. Use for any question about past conversations, previous topics, what the user asked before, or recent activity. Replaces the memory skill.
| name | memory |
| description | Retrieve past conversations. Use for prior dialogue, past questions, or history lookup. |
| metadata | {"type":"primitive","dependencies":"time"} |
Short-term first, long-term second: For recall ("what did we discuss about X"), prefer recent context. memory.search automatically queries current session first, then expands to cross-session. memory.read defaults to current session when date and session_id omitted.
Use search for keyword lookup; use read for a specific date or session. Omit session_id to use current session (short-term).
Resolving referents: When the user refers to "my question", "what I asked", "the previous message", "that content", etc. without stating it explicitly, prioritize the most recent rounds of dialogue (default 3 rounds; configurable via referent_context_rounds). Only the most recent N rounds are passed to you, so use them first. Call memory.read (omit session_id for current session) if you need more. If you cannot determine the referent confidently, ask the user to clarify before acting.
For relative dates (yesterday, 2 days ago, day before yesterday): call time.calculate first. Use its since/until (extract YYYY-MM-DD) as memory.search(date_range={"start": since_date, "end": until_date}) or memory.read(date=since_date).
Search memory by keywords. Short-term (current session) first, then long-term (cross-session).
Read complete memory. Defaults to current session (short-term) when date and session_id omitted.
Summarize memory over a period.
