researcher

Analyze a paper from a reviewer's perspective — find problems, weak claims, research gaps. Stands alone (no orchestrator wraps it). Triggers: 问题分析, 找论文弱点, 审稿式分析, critic, reviewer critique, 审稿质疑.

Design small, controllable validation experiments for a paper's identified problems (from critic's output). Triggers: 验证实验设计, design validation experiments, 设计验证实验.

Generate or persist evidence-grounded research idea cards anchored to named pain points. Use for research ideas, research directions, or a brainstorm workflow that needs candidate cards or reviewed-card persistence.

Curate the research wiki — lint, compare papers, answer literature queries. Triggers: 文献查询, 对比论文, 跨论文比较, wiki lint, 文献检索, literature query, compare papers, cross-paper.

Deeply extract a paper's experiments for downstream critique and validation design. Triggers: 实验提取, 深度提取, deep extraction.

Unified wiki write entry — ingest a paper PDF into an evidence-traceable paper page, OR write an already-prepared markdown file into the wiki. Triggers: 入库, ingest paper, add to wiki, 写入 wiki, 把这份文档入库, 存进 wiki.

Query and compare literature across wiki papers — runs the curate predicate. Triggers: 文献查询, 对比论文, 跨论文比较, wiki里有没有, literature query, compare papers, cross-paper.

Deeply read a paper in one flow — ingest into the wiki, then extract its experiments (ingest → extract). Triggers: 深读论文, 精读这篇, paper-read, deep read.

Turn a paper's problems into executable validation (design → spec). Requires critic's analysis already in the wiki. Triggers: 验证设计, paper-validate, 验证实验, 怎么验证这个 claim.

Send a structured inter-session message via sessions_send — single source of truth for the call format, XML message wrapping, and field spec. Triggers: 发送会话消息, 回传结果, 发给上游, sessions_send, 报告 blocker.

spec — Translate validation designs (from design's output) into a claude-code task prompt. Triggers: 生成任务提示词, spec, codex 提示词, claude-code task, 实现规格.

Audit the quality of analysis-chain outputs (extract/critic/design/spec) — read-only. Triggers: 质量审计, audit, 质量评估, 流水线审计.

Run parallel, evidence-grounded research ideation with main-agent rebuttal. Use for research brainstorming, research directions, or finding research gaps from a paper/wiki context.

Quality-gate audit of a paper's analysis outputs (extract/critic/design/spec) — runs the audit predicate. Triggers: 质量审计, paper-audit, 审计产出, 流水线审计, 检查这些产出.

Ingest a paper into the wiki and verify quality (ingest → curate). Triggers: 入库, add to wiki, 文献笔记, 整理这篇论文.

Coordinate inter-session communication — when and why to send messages, how to structure spawned tasks. Message format lives in `send`. Use when a caller must pass a callback session key and full skill prompt to a callee, a callee must promptly report a blocker or material finding, or a task needs inter-session updates before final delivery.

Batch-ingest many papers - trigger the ingest skill once per paper. Triggers: 批量入库, 多篇论文录入, 批量录入论文, batch ingest, ingest multiple papers.

Create, validate, update, and sync OpenClaw agent workspaces. Handles the full lifecycle: scaffolding new workspaces with all required files (SOUL.md, AGENTS.md, IDENTITY.md, USER.md, TOOLS.md, MEMORY.md, HEARTBEAT.md, DREAMS.md), updating individual workspace files while preserving existing conventions, validating consistency between workspace directories and openclaw.json registration, and syncing agent configs. Use this skill whenever the user mentions workspaces, agent setup, subagent creation, workspace files (SOUL/AGENTS/IDENTITY/etc), workspace validation, or any task that involves creating or modifying agent workspace directories. Also use when the user says "add a new agent", "set up workspace", "check workspaces", "fix workspace", "sync agent config", or mentions workspace-maintenance tasks — even if they don't explicitly say "workspace-doctor".

Andrej Karpathy的思维框架与表达方式。基于30+个一手来源（karpathy.ai、GitHub、YouTube、Twitter）的深度调研， 提炼5个核心心智模型、7条决策启发式和完整的表达DNA。 用途：作为AGENTS.md的系统提示词审阅顾问，用Karpathy的视角分析agent指令设计、文档质量、抽象边界。 触发：当用户提到「用Karpathy的视角」「Karpathy会怎么看」「karpathy模式」「karpathy perspective」「Karpathy review」时使用。 即使用户只是说「帮我看下这个prompt」「这个AGENTS.md写得怎么样」并提到Karpathy，也应触发。

女娲造人：输入人名/主题/甚至只是模糊需求，自动深度调研→思维框架提炼→生成可运行的人物Skill。 两种入口：(1)明确人名→直接蒸馏 (2)模糊需求→诊断推荐→再蒸馏。 触发词：「造skill」「蒸馏XX」「女娲」「造人」「XX的思维方式」「做个XX视角」「更新XX的skill」。 模糊需求也触发：「我想提升决策质量」「有没有一种思维方式能帮我...」「我需要一个思维顾问」。

Complete OpenClaw agent and configuration management. Use when the user wants to create new agents, configure Telegram bots, set up message routing with bindings, manage multi-agent configurations, list current setup, or troubleshoot routing issues. Handles agent creation, binding management, and configuration based on official OpenClaw documentation.