菜单
Forces the agent to actively research official documentation, GitHub issues, and installed package versions before implementing new features OR debugging failures.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
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。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/misunders2d/ori --skill external-research-skill命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Working memory protocol for multi-step tasks. Use the scratchpad to record intermediate findings without bloating LLM context.
How to schedule one-off and recurring agent tasks, route their delivery to specific Slack/Telegram channels, and use plan enforcement reliably. Load this any time the user asks to remind, schedule, automate, or run something on a cron.
How to add, list, or remove API keys and OAuth integrations through the chat. Load this when the user asks to add/set/configure/connect a key, token, credential, or third-party service (Slack, Telegram, OpenRouter, Anthropic, Google, GitHub, ClickUp, …).
基于 SOC 职业分类
正在显示 SKILL.md
| name | external-research-skill |
| description | Forces the agent to actively research official documentation, GitHub issues, and installed package versions before implementing new features OR debugging failures. |
Your pre-training knowledge has a cutoff date. Libraries change APIs, deprecate features, and introduce breaking changes. You MUST NOT guess or hallucinate implementations.
check_installed_package — confirm what's actually installed locally.google_search_agent_tool — locate official documentation.web_fetch — fetch docs (converts HTML to markdown, ~90% token reduction).You get ONE attempt from your own knowledge. If verification fails, you MUST research externally before retrying. Every attempt after the first must be backed by external evidence.
Read references/research-best-practices.md for detailed guidance on source credibility scoring and RRV orchestration.
web_fetch returns markdown, not HTML: Don't try to parse HTML tags from its output. The conversion is automatic.llms.txt: If a target repo provides this file at the root, fetch it first — it's a structured documentation index purpose-built for LLMs. Saves tokens vs. crawling multiple pages.robots.txt and agents.txt: Respect them. If a site blocks agents, don't try workarounds.web_fetch returns empty/boilerplate content, the site is client-rendered. Note this and move on — don't burn retries.check_installed_package before assuming an API exists. A function that works in v2.0 may not exist in v1.8.