AI at AquaGen

Before you start

Pick the tool by what you want to end up with. Every tool here sends your prompts and files to the provider, so the AquaGen AI policy applies to all of them.

Which tool for which job

Use it with judgement

• Data leaves AquaGen. Prompts and files are processed by the provider and stored in the US. Follow the AquaGen AI policy, and do not share sensitive data without approval.
• Public services see your input. Some skills send data to public servers directly. Structure prediction, for example, sends protein sequences to the public ColabFold server.
• You own the conclusion. Check the code, the sources and the numbers before you use a result. A built-in reviewer is a second pair of eyes, not validation.
• Passwords stay with you. Never type a password, key or token into a chat. Type it only in your own terminal.

Access

ChatGPT: sign in with your AquaGen work email. Skills depend on workspace settings, so some options may be hidden until an admin enables them.

Claude Science: needs a paid Claude plan and is not available to everyone. If you want to be considered for access, contact IT.

ChatGPT on your PC, working on our cluster

Install the ChatGPT app on Windows and connect it to the cluster over SSH. Chats in a project folder then run on the cluster, next to the data and the compute.

You need: Windows 10 or 11, a user account on the cluster, and your AquaGen work email. Plan for about 15 minutes.

Watch first, then follow the stepsAn 88-second video with the whole flow, plus the written steps, manual commands and troubleshooting table. Screens in the video are simplified, and host names are placeholders.timknutsen.no/gpt-install

1. Install from the Microsoft Store

   Search for ChatGPT and pick the app published by OpenAI.

2. Open and log in

   Sign-in finishes in your browser. Use your work email.

3. Add the cluster and switch it on

   Settings › Connections › SSH › Add. Address: the cluster address you got from IT. On Norwegian Windows the labels are Koding, Tilkoblinger and Legg til.

   The first connection usually fails with “SSH connection failed”. That is expected: passwordless login is not set up yet.

   Shortcut: one prompt for steps 4 to 6. Start a local chat (one that runs on your PC, not on the host), replace CLUSTER-ADDRESS and YOUR-USERNAME with the cluster address and your username, and paste:

   I'm on Windows and want the ChatGPT app to work on our cluster over SSH. Host: CLUSTER-ADDRESS, alias cluster, my username: YOUR-USERNAME. Please do this in order, and never ask me for my password: 1. Check my SSH setup in %USERPROFILE%\.ssh. If I have no key, create an ed25519 key. Add or fix a "Host cluster" entry in my SSH config with the address, my username and the key. 2. Give me one script that adds my public key to ~/.ssh/authorized_keys on the server with the right permissions (700 for ~/.ssh, 600 for authorized_keys). I will log in with my password in my own PowerShell window, run the script there, and type done here. 3. When I type done, test "ssh cluster" without a password and fix anything that still fails. 4. Over SSH, check whether Codex CLI is installed and on the PATH for non-interactive logins (codex --version). If not, install it with npm install -g @openai/codex, using a user-level npm prefix in my home folder if I lack admin rights, and add it to the PATH in ~/.bashrc and ~/.profile. 5. Confirm codex --version works over a fresh SSH login, then tell me to reconnect cluster under Settings › Connections › SSH.

   Copy

   The same prompt also repairs a host that stopped working, for example after a new PC or a new key. Steps 4 to 6 below do the same thing one piece at a time.

4. Ask a local chat to fix the login

   Start a chat that runs on your PC, not on the host, and paste:

   Fix passwordless SSH login to cluster and give me the script to run on the server

   Copy

   Codex checks your SSH setup, creates a key if the PC has none, and writes a short script that adds the key on the server.

5. Run the script yourself

   Open a separate PowerShell window, ssh to the server and type your password there. Paste the script, run it, then type done in the chat. Codex tests the login again.

6. Install Codex CLI on the server

   If the app says Codex CLI is not installed on the remote machine, ask the same chat:

   Codex CLI isn't installed on cluster. Install it over SSH now.

   Copy

7. Reconnect and pick a folder

   Reconnect the host and sign in with your work email. Choose a project folder on the cluster. Chats in that folder now run on the cluster.

When it fails

ChatGPT skills: teach it a workflow once

A skill is a saved set of instructions, examples and optional scripts for one task, so ChatGPT does it the same way every time.

Where skills live and how they are used

• Find them: Sidebar › Plugins › Skills. Tabs: Installed, Created by me, Shared with me, Shared by AquaGen.
• Use them: ChatGPT picks a skill when your task matches its description. You can also name it with @.
• Availability: Business, Enterprise, Healthcare and Edu workspaces, subject to workspace settings.

A skill is a folder

Skills follow the open agent skills standard: a folder with a SKILL.md file holding a name, a description and the instructions. The description decides when the skill is used, so write it as the situation that should trigger it.

genotype-qc/
  SKILL.md
    name: genotype-qc
    description: standard QC for salmon PLINK files
    steps, checks, report
  scripts/      (optional)

Illustrative example, not an existing AquaGen skill.

Create or install a skill

1. Create with chat (quickest)

   Ask in a normal chat. The built-in skill-creator asks what the skill should do and when it should trigger, then offers to install it.

   Make a skill for our standard genotype QC.

   Copy

2. Create with editor

   Skills › Create › Create with editor, and write the instructions yourself.

3. Upload

   Skills › Create › Upload from your computer. ChatGPT scans the upload and may mark it Needs Review or Blocked. Review any skill from an outside source before you upload it.

4. Install a shared one

   Shared with me or Shared by AquaGen › ••• › Install.

Claude Science

A desktop app for research analysis. You describe the task in plain language; Claude writes and runs Python, R or shell code on your computer or on our cluster, and saves every result with the code and conversation that made it.

Get started

1. Install

   Download the installer from claude.com/product/claude-science and open it. It installs for your user account and adds Claude Science to the Start menu. You need Windows 11 (or macOS 13 or later) and about 5 GB of free disk space.

   The first start is slow: it downloads about 150 MB and sets up Python and R for a few minutes.

2. Sign in

   Click Sign in on the web, then Continue, and approve in your browser. No API key is needed.

3. Go through the setup wizard

   It asks which connectors and skills to switch on, which websites Claude may reach, and whether memory is on. The defaults are fine; everything can be changed later under Settings.

4. Run a first analysis

   Open the Example project or create a new one. In the message box, type @ to point Claude at a folder on your PC, then describe the task.

   Claude asks before it reads a new folder and before it runs code. Results appear as artifacts in the Files panel.

Connect our cluster (optional)

Heavy jobs can run on the cluster. Claude Science uses your existing SSH setup, so do the ChatGPT shortcut prompt above first; it creates the key and the cluster entry that Claude Science reads.

1. Add the host

   Settings › Compute › SSH hosts › Add SSH host. Pick cluster from the list, optionally add notes (partition, modules, where data lives), and click Add. Claude probes the host and saves what it finds.

2. Approve jobs

   When Claude wants to run a job, a card shows the command and script. Approve it once, for the conversation, or for the project. Jobs go to Slurm and keep running if you lose the connection.

Remote jobs run as your own user, with access to everything your account can read and write on the cluster. The default time limit is 30 minutes, so tell Claude if a job will run longer.

Good to know

• Databases: connectors give read-only access to public sources such as Ensembl, UniProt, PubMed and bioRxiv. Switch them on under Settings › Connectors.
• Methods: skills hold ready-made methods, from literature review to AlphaFold2. Type / to pick one, or just describe the task.
• Checks: a background reviewer checks Claude's claims against the code that actually ran. It reduces errors but does not remove them, so verify before you rely on a result.

Full documentation: claude.com/docs/claude-science.

Claude plugins

A plugin bundles skills, connectors, commands and agents. Add it once in your Claude account and it works in chat, Cowork and Claude Code.

1. Find

   claude.ai › Customize › Plugins › Discover.

2. Add

   Read what it contains, then Add.

3. Connect

   Adding a plugin signs you in to nothing. Open its Connectors tab and connect each service you need.

4. Use

   Type / and pick plugin-name:skill-name, or describe the task.

Example: the Bio Research plugin

Try it on our questions

These prompts come from real Claude Science sessions. Swap in your own sequence, species or question, and check every result.

Map a sequence on the cluster

Take stock of public data

Search the literature

Plan an experiment

One task this week

1. Pick a task that takes too long

   Something you repeat, or have put off.

2. Choose the tool

   ChatGPT on the cluster, or Claude Science.

3. Check it and share what worked

   Bring it to the next AI Champions session. Ask Tim Knutsen for help with setup.

Sources