Liu Lab Skills¶

Agent Skills that teach AI coding agents how the Liu Lab works, starting with its HPC clusters. Built for Claude Code, and they work in any tool that reads the open skills standard.

Ask an agent to "run this on the cluster" and it needs lab knowledge it cannot guess: which clusters exist, how a job gets submitted, what is safe to do. This plugin holds that knowledge.

Skill	What it covers
lab-hpc	The two clusters (arc/chimera GPU, ircbc CPU), Slurm, the safety rules, the per-machine setup check.
lab-jupyter	Jupyter Lab on a compute node, tunnelled to your local browser.
lab-containers	The lab's Singularity images on ircbc: keeping them current, running work inside them.
lab-edison	The Edison platform: Kosmos for a whole research goal, cited literature answers and data analysis for a single question. No cluster.
lab-setup	One step after installing: the plugin updates by itself, and Codex and Kimi Code get it too.
hq-mode	A whole task done the lab's way: working rules, a playbook per kind of task, and a list of steps that always wait for your yes.
lab-unslop-zh	The last pass over Chinese writing: it writes Chinese rather than translating it, and leaves your voice and your facts alone.
lab-unslop-zh-kimi	A Kimi candidate for natural Chinese writing and strict technical translation, with protected mixed content.

The three cluster skills load by themselves when your request touches remote compute. The Chinese writing skills load when the deliverable is Chinese, and you can also call them by name. Call lab-edison, lab-setup and hq-mode by name: lab-edison because every run costs you money, lab-setup because you need it once per machine, and hq-mode because you choose when a task gets it.

What the agent will not do¶

It never runs compute on a login node, never submits a job you have not seen, and never picks a cluster in silence. On a machine that is not set up it refuses and points you at the setup below, rather than asking for an address or a password. When ircbc stops answering it tells you to check the VPN instead of retrying in a loop.

No secrets in this repo

The skills hold no hostnames, addresses, usernames or keys. Hosts appear only as ssh alias names, and the rest is read from your ~/.ssh/config at run time. That is what makes this repo safe to publish.

Install¶

Two commands in your terminal:

claude plugin marketplace add liuhlab/liulab-skills
claude plugin install lab@liulab

Then start Claude Code and type /lab-setup. It turns on updates for the plugin, and sets up Codex and Kimi Code if you have them. It also points each agent at the lab's defaults: where issues go, which labels to use, and the safety rules for the clusters. It asks before each change, and it removes nothing. Its page has the details.

Then set the machine up for the clusters once.

SSH aliases. Your ~/.ssh/config needs the lab's usual alias names: arc, chimera-login, chimera-transfer, chimera-gpu, chimera-cpu, the per-node aliases, ircbc, ircbc-transfer, and cpu01 through cpu08, each with your own username and key. Ask a lab member for a working example. This repo ships none.
Personal config. Copy templates/personal.md to ~/.claude/lab/personal.md and fill in your usernames, directories and Jupyter defaults. Never commit it.
Edison key and pixi, only for lab-edison. That skill needs pixi on your PATH. Then copy templates/edison.env to ~/.claude/lab/edison.env, paste your own key over the placeholder, and chmod 600 it. It sits beside personal.md, never inside it. Never commit it.

To check the result, ask Claude: "Am I set up for the lab HPC clusters?" The skill runs its preflight and reports per cluster.

Codex and Kimi Code¶

/lab-setup sets both up from Claude Code, so there is nothing to install by hand. Codex gets the same plugin and checks for a new release each time it starts. Kimi Code reads the skills from the copy Claude Code keeps up to date. No other agent tool is supported.

lab-edison waits for you to type its name in all three. Claude Code and Kimi Code read that from its settings. Codex reads it from a small opt-out file the skill also ships, agents/openai.yaml.

Updating¶

Once /lab-setup has run, there is nothing to do: each agent picks up a new release on its own. To update by hand:

claude plugin marketplace update liulab
claude plugin update lab@liulab

Claude Code loads plugins when it starts, so restart your session before the updated skills are there.

Moving from `lab-compute`¶

Until 2026.10.1 this repo was liulab-compute-skills and the plugin was lab-compute. A plain update does not cross a rename, so do it by hand once:

claude plugin uninstall lab-compute@liulab
claude plugin marketplace remove liulab
claude plugin marketplace add liuhlab/liulab-skills
claude plugin install lab@liulab
mv ~/.claude/compute ~/.claude/lab

The last line moves your own files to where the skills now look for them. Every skill refuses to run while ~/.claude/compute is still there. If you installed from a clone instead, use the two commands under Install. /lab-setup then lists the links your old install left, for you to remove.