Kalaris Labs开源281个SKILL.md工作流解决Agent幻觉
[Open Source / MIT] How we stopped coding agents from hallucinating citations and repeating failed ML experiments (281 modular SKILL.md workflows
做AI Agent开发的同学必看,这套SKILL.md方案直接解决了引用幻觉和实验状态丢失两大顽疾,流程清晰可复用,建议直接参考其验证逻辑优化你的Agent链路。
[Disclosure per Rule 5 & Rule 10: I am an AI researcher at Kalaris Labs contributing to this repository alongside creator Sayan Chowdhury. It is 100% free and FOSS-licensed (MIT) with no commercial tier, no locked "pro" features, and no paywall. Written by a human researcher.]
[根据规则5和规则10的披露:我是Kalaris Labs的AI研究员,与仓库创建者Sayan Chowdhury共同贡献于此仓库。该项目完全免费且采用FOSS许可(MIT),无商业层级、无锁定“专业”功能、无付费墙。由人类研究员撰写。]
While working on LLM research papers, our team kept running into three failure modes with coding and research agents:
在撰写LLM研究论文期间,我们的团队在与编码和研究智能体协作时反复遇到三种故障模式:
- Corrupted references in .bib files: LLMs frequently pair a real paper title with the wrong publication year, invent co-authors, or cite a DOI that resolves to an unrelated work. Simply checking HTTP 200 on a DOI link misses metadata mismatches and retractions.
- Stateless experiment loops: When an agent runs a batch of ML experiments across sessions, context compaction wipes out earlier negative results, causing the agent to re-test variations of failed hyperparameters.
- Context bloat from monolithic system prompts: Loading dozens of tool definitions and domain rules into every session degrades instruction adherence.
- .bib文件中的引用损坏:LLM经常将真实的论文标题与错误的出版年份配对,捏造合著者,或引用指向无关作品的DOI。仅检查DOI链接的HTTP 200状态会遗漏元数据不匹配和撤稿情况。
- 无状态实验循环:当智能体跨会话运行一批ML实验时,上下文压缩会清除早期的负面结果,导致智能体重试已失败超参数的变体。
- 来自单体系统提示词的上下文膨胀:将数十个工具定义和领域规则加载到每个会话中会降低指令遵循能力。
We open-sourced research-agent-skills (https://github.com/KalarisLabs/research-agent-skills), an MIT-licensed collection of 281 domain-specific skills for Claude Code, Codex, Cursor, OpenCode, Gemini CLI, and Windsurf.
我们开源了research-agent-skills(https://github.com/KalarisLabs/research-agent-skills),这是一个MIT许可的集合,包含针对Claude Code、Codex、Cursor、OpenCode、Gemini CLI和Windsurf的281个特定领域技能。
How the verification and routing architecture works (with sources per Rule 4):
验证和路由架构的工作原理(依据规则4提供来源):
- Selective per-project installation: Running npx skills add KalarisLabs/research-agent-skills --skill <name> installs only the specific SKILL.md and helper scripts needed for that repo, avoiding context window bloat (see architecture docs: https://docs.kalarislabs.com/research-agent-skills).
- Deterministic citation verification (citation-verification): Instead of asking the LLM to judge whether a citation is real from its weights, the skill executes a zero-dependency Python 3.9+ standard library script against Crossref (https://api.crossref.org), OpenAlex (https://api.openalex.org), arXiv (https://export.arxiv.org), and DataCite (https://api.datacite.org), plus Retraction Watch relation metadata in Crossref. Every .bib entry is classified into verified, found-add-doi, mismatch, doi-not-found, uncertain, or not-found. The agent is only allowed to apply mechanical DOI additions (found-add-doi) and must pause for human confirmation on any mismatch or not-found entry.
- Two-loop experiment orchestration (autoresearch): Splits autonomous experimentation into an inner loop (rapid experiment iterations against a target metric) and an outer loop that maintains findings.md. Hypotheses must state a testable mechanism ("X because Y, predicting Z"), and negative results are logged with what they rule out.
- Prompt-injection & trigger testing: Every skill is evaluated against reviewed prompt sets for trigger routing accuracy and scanned offline for prompt injection using Skill Doctor (https://docs.kalarislabs.com/).
- 按项目选择性安装:运行 npx skills add KalarisLabs/research-agent-skills --skill <name> 仅安装该仓库所需的特定SKILL.md和辅助脚本,避免上下文窗口膨胀(参见架构文档:https://docs.kalarislabs.com/research-agent-skills)。
- 确定性引用验证(citation-verification):与其让LLM根据其权重判断引用是否真实,该技能会执行一个基于Python 3.9+标准库的零依赖脚本,对接Crossref(https://api.crossref.org)、OpenAlex(https://api.openalex.org)、arXiv(https://export.arxiv.org)和DataCite(https://api.datacite.org),以及Crossref中的Retraction Watch关系元数据。每条.bib条目被分类为verified(已验证)、found-add-doi(找到并添加DOI)、mismatch(不匹配)、doi-not-found(DOI未找到)、uncertain(不确定)或not-found(未找到)。智能体仅允许应用机械式的DOI添加(found-add-doi),并在任何不匹配或未找到的条目上暂停以等待人工确认。
- 双环实验编排(自动研究):将自主实验拆分为内环(针对目标指标的快速实验迭代)和外环(维护 findings.md)。假设必须声明可测试的机制(“X 因为 Y,预测 Z”),并将阴性结果记录为排除了什么。
- 提示注入与触发器测试:每项技能均使用审核过的提示集评估触发路由准确性,并使用 Skill Doctor(https://docs.kalarislabs.com/)进行离线扫描以检测提示注入。
GitHub repository (MIT): https://github.com/KalarisLabs/research-agent-skills
GitHub 仓库(MIT 许可证):https://github.com/KalarisLabs/research-agent-skills
Full 281-skill catalog: https://docs.kalarislabs.com/research-agent-skills/skills
完整 281 项技能目录:https://docs.kalarislabs.com/research-agent-skills/skills
We are currently expanding our ML evaluation and mechanistic interpretability modules. If you build agentic workflows for research or technical writing, I would appreciate public feedback in the comments on where this verification pipeline could fail or which scholarly APIs we should add next.
我们目前正在扩展机器学习评估和机械可解释性模块。如果您为研究或技术写作构建智能体工作流,我将不胜感激您在评论中提供公开反馈,指出此验证流程可能失效之处,以及我们应该添加哪些学术 API。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力