Reusable Agent Skills distilled from papers and open-source repos, covering AI agent engineering (authorization, tooling, workflows) and research / academic workflows (papers, citations, figures, literature). Each skill is a folder with a SKILL.md (YAML frontmatter + recipe) and an examples.md (runnable / follow-along usage examples with acceptance checks).
| Skill | Source | One-line purpose |
|---|---|---|
| sara-tool-authorization | arXiv:2608.27146 | Stop untrusted tool outputs (email, web, MCP results) from silently becoming permission to act. |
| openai-agents-sdk-sandbox | openai/openai-agents-python | Build multi-agent workflows with the OpenAI Agents SDK: SandboxAgent, MCP, guardrails, handoffs, Realtime. |
| state-path-tool-menus | arXiv:2609.09395 | For large tool/MCP libraries, build an ordered, executable tool menu (producers before consumers) instead of relevance-only retrieval. |
| Skill | Source | One-line purpose |
|---|---|---|
| acl-latex-paper | Iris / TalentOS | Write/compile ACL-family LaTeX papers (template, anonymity, page limits, compile checks). |
| acl-paper-figures | alphaXiv/OpenResearch orx-figures (MIT, adapted) | Build print-size ACL/two-column matplotlib figures (Okabe-Ito, greyscale-safe) with audit helpers. |
| annotation-project | Iris / TalentOS | Design labeled datasets: scheme, guidelines, IAA, LLM pre-annotation, data statement/licensing. |
| citation-audit | wanshuiyin/Auto-claude-code-research-in-sleep (MIT, adapted) | Audit every bibliography entry for reality, attribution, and claim support (MBA/CSSCI-aware). |
| citation-verification | Iris / TalentOS | Resolve every citation against DOI/Crossref/Anthology/arXiv before sharing refs. |
| deslop | stephenturner/skill-deslop (MIT, adapted) | Strip AI writing tells and defensive hedging from academic prose (EN + Chinese journal extras). |
| literature-review | Iris / TalentOS | Systematic literature search protocol, multi-source querying, snowballing, review matrix. |
| openalex | wanshuiyin/Auto-claude-code-research-in-sleep (MIT, adapted) | Search OpenAlex for works, citation counts, affiliations, funding, DOI metadata via public API. |
| rebuttal-letter | Iris / TalentOS | Structure conference rebuttals / journal response-to-reviewers letters. |
| research-skill-router | Iris / TalentOS | Route research tasks to the right local skill; install external MIT/Apache skills on demand. |
| scientific-visualization | K-Dense-AI (MIT) | Post-figure QA: palette/contrast audit, PDF metadata, export planning. |
| scipilot-figure-skill | Haojae/scipilot-figure-skill (MIT) | Chart-type advice and viz anti-pattern checklist. |
| semantic-scholar | wanshuiyin/Auto-claude-code-research-in-sleep (MIT, adapted) | Search Semantic Scholar for published venue papers with citations/TLDR via public API. |
| venue-deadlines | Iris / TalentOS | Track official CFP deadlines (AoE conversion) and sourced acceptance rates. |
Copy a skill folder into your agent's skills directory, for example:
git clone https://github.com/TalentOS-AI/ai-agent-skills.git
cp -r ai-agent-skills/skills/sara-tool-authorization <your-agent>/skills/The description field in the frontmatter is what the agent uses to decide when to apply the skill. Agents that have no skills directory can take SKILL.md pasted into their instructions.
.
├── LICENSE
├── README.md
├── templates/
│ └── SKILL.template.md # starting point for new skills
└── skills/
└── <skill-name>/
├── SKILL.md # the skill itself
├── examples.md # usage examples + acceptance checks (engineering skills)
└── … # research skills may ship helpers/scripts/references instead
- Read the paper or repo thoroughly (problem, method, results, limits).
- Copy
templates/SKILL.template.mdtoskills/<kebab-name>/SKILL.mdand fill it in. - Add
examples.mdwith at least 2 examples (input prompt, expected behavior, acceptance checks). - Link the original source. Never invent citations, benchmark numbers, or star counts; only use numbers you verified at the source and say where they came from.
- Open a pull request.
MIT. Skill text is free to reuse; rights to the underlying papers and repos stay with their authors. Always cite the source.
从论文和开源仓库提炼出的可复用 Agent Skill,覆盖 AI Agent 实操(授权、工具、工作流)与科研/学术流程(论文、引用、作图、文献)。每个 skill 是一个文件夹,内含 SKILL.md(YAML 头 + 做法)和 examples.md(可运行或可照做的示例与验收要点)。
| Skill | 来源 | 一句话用途 |
|---|---|---|
| sara-tool-authorization | arXiv:2608.27146 | 防止邮件、网页、MCP 返回等不可信工具输出悄悄变成执行授权。 |
| openai-agents-sdk-sandbox | openai/openai-agents-python | 用 OpenAI Agents SDK 搭多 Agent 工作流:SandboxAgent、MCP、guardrail、handoff、Realtime。 |
| state-path-tool-menus | arXiv:2609.09395 | 工具/MCP 很多时,按可执行路径排工具菜单(生产者在前、消费者在后),而不是只按相关度检索。 |
| Skill | 来源 | 一句话用途 |
|---|---|---|
| acl-latex-paper | Iris / TalentOS | 撰写与编译 ACL 系 LaTeX 论文(模板、匿名、页限、编译检查)。 |
| acl-paper-figures | alphaXiv/OpenResearch orx-figures(MIT,改编) | 按 ACL/双栏印刷尺寸做 matplotlib 图(Okabe-Ito、灰度安全)并自检。 |
| annotation-project | Iris / TalentOS | 标注数据集:方案、指南、一致性、LLM 预标、数据说明与许可。 |
| citation-audit | wanshuiyin/Auto-claude-code-research-in-sleep(MIT,改编) | 逐条核验参考文献是否真实、归因正确、支撑主张(兼顾 MBA/CSSCI)。 |
| citation-verification | Iris / TalentOS | 分享参考文献前用 DOI/Crossref/Anthology/arXiv 解析每条引用。 |
| deslop | stephenturner/skill-deslop(MIT,改编) | 去掉学术文风里的 AI 套话与防御性表述(含中文期刊补充)。 |
| literature-review | Iris / TalentOS | 系统文献检索:多源查询、滚雪球、综述矩阵。 |
| openalex | wanshuiyin/Auto-claude-code-research-in-sleep(MIT,改编) | 用 OpenAlex 公开 API 查论文、引用、机构、资助与 DOI 元数据。 |
| rebuttal-letter | Iris / TalentOS | 组织会议 rebuttal / 期刊回复审稿意见信。 |
| research-skill-router | Iris / TalentOS | 把科研任务路由到合适本地 skill;按需安装外部 MIT/Apache skill。 |
| scientific-visualization | K-Dense-AI(MIT) | 成图后质检:配色对比、PDF 元数据、导出规划。 |
| scipilot-figure-skill | Haojae/scipilot-figure-skill(MIT) | 图表类型建议与可视化反模式清单。 |
| semantic-scholar | wanshuiyin/Auto-claude-code-research-in-sleep(MIT,改编) | 用 Semantic Scholar 公开 API 查已发表会议/期刊论文与引用/TLDR。 |
| venue-deadlines | Iris / TalentOS | 跟踪官方 CFP 截稿(AoE 换算)与有出处的录用率。 |
把 skill 文件夹复制到 Agent 的 skills 目录(命令见上方英文部分)。Agent 依据 frontmatter 里的 description 判断何时启用;没有 skills 目录的 Agent,可以把 SKILL.md 内容粘进指令。
见上方英文部分:templates/ 放模板,skills/<名称>/ 下是 SKILL.md 和 examples.md。
- 完整读懂论文或仓库(问题、方法、结果、局限)。
- 复制
templates/SKILL.template.md到skills/<kebab-name>/SKILL.md并填写。 - 写
examples.md,至少 2 个示例(输入提示、预期行为、验收要点)。 - 附原始出处。不编造引用、基准数字或 star 数,数字必须在来源处核实并注明出处。
- 提交 Pull Request。
MIT。Skill 文本可自由复用,论文与仓库的权利归原作者,使用时请注明来源。
Maintained by TalentOS · 需要定制 Agent / 技能开发?找我们定制 / Work with us · talentos@agentbay.asia