What Is Claude Science?

What Is Claude Science?

Part of the playlist Northeastern HPC, AICR and Claude Science. A tour of a real session of Claude Science, Anthropic's app for research analysis: what the app is, what it asks before it runs anything, how it keeps a record of every result, and what it remembers and counts.

Everything on screen is a real capture or recording of the app, or real command output from one Mac. Nothing is mocked. Counts such as the number of skills and connectors are from the beta we tested and will change. The film does not show a connection to a cluster.

0:00 Introduction 0:43 The app itself 1:17 It asks before it acts 2:05 Every result has a history 2:57 It reaches real data 3:57 It remembers and it counts 4:40 Recap 4:58 Your turn

Your turn: read one small public data file from a folder you grant Claude Science read-only, fit a simple trend, save a figure as an artifact named trend.png, then redraw it with a logarithmic x axis under the same name. Check that every code card showed the code before it ran, and that the figure opens with two versions you can step between.

Documentation: claude.com/docs/claude-science

youtube.com/@NikBearBrown

▶ The full series playlist — Brutalist https://www.youtube.com/playlist?list=PLG9H-C6rp5RU

What this video covers

• 0:00 — Introduction
• 0:43 — The app itself
• 1:17 — It asks before it acts
• 2:05 — Every result has a history
• 2:57 — It reaches real data
• 3:57 — It remembers and it counts
• 4:40 — Recap
• 4:58 — Your turn

Transcript

A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.

Introduction

[0:00] Northeastern HPC, AICR, and Claude Science. What is Claude Science? A tour of a real session. What the app is, what it asks before it runs anything, and how it keeps a record of every result. Hello, this is Liam in for Bear. First, a correction. Claude Science is not a chatbot that happens to know biology. It is a workspace on your own computer where Claude runs real code, asks first, and keeps the receipts. Six terms, an artifact, a file Claude saves with its versions. Provenance, the record of how a result was made. A sandbox, a fenced area where Claude's code runs. A permission card, a question Claude asks before it acts. A connector, a link to an outside database or tool. And a skill, a saved set of instructions for one kind of task. Claude Science is

The app itself

[0:43] A program on your own computer. Ask the terminal where it lives and it answers. Ask its version and it answers. Ask its status and you see it is running, the port it serves, that its sandbox is on, and that skipping approvals is off. That port serves a web page and the page is the app. Projects on the left, recent sessions on the right. A project groups related work and each session is one conversation inside it. Each project carries its own instructions and Claude reads them in every session. Our say, use only public data and keep answers short. Typed once, applied every time. Now we ask it to read a file and fit a

It asks before it acts

[1:18] Curve. Before it runs a single command, a card appears with the exact code. Nothing runs until you allow it or deny it. Allow has a scope once, this conversation, this project, or everywhere. The default is the whole conversation. We picked once every time, so every single command had to ask again. Files are asked for separately. Claude needed one folder, so it asked for that one folder, read only or read and write. We chose read only. Code can fail and the page says so. Here are three failed attempts in a row, a missing match, a missing name, a wrong type. Claude read each error, fixed it, and each fix came back as a new card to approve. Allowing a folder is remembered. Settings permissions now lists that folder read-only with a global scope. A command allowed once leaves no entry at all.

Every result has a history

[2:05] The figure is saved as an artifact and every save is a version. Here is version two with a log axis. Step back and version one appears, the plain axis. Same name with a history. Open its provenance and there are five tabs: code, execution log, messages, environment, review. The code tab carries a warning. It is a reconstruction written by Claude. The execution log is the raw record. The environment tab pins the exact Python and every package, 121 of them, each with its version. That is what lets someone rebuild the result later. In this session, after each reply, a second model, the reviewer, read the work. It flagged one thing, a link written the wrong way, a formatting slip, not a wrong answer. It marked it a warning still unaddressed. It ran by itself because auto review was on, a switch in the session's options. Delegation and memory were on beside it.

It reaches real data

[2:57] Type a slash and skills open. A skill is a saved set of instructions for one kind of task: protein structure, single-cell data, literature review. In the beta we tested, the list held 32. Connectors reach outside databases. Settings lists 36 and 20 of them are research databases from Anthropic, each marked connected. We asked for human hemoglobin in UniProt and the Protein Data Bank. Claude listed its attached connectors, found none for those two, and said so. Then it took another route, code that calls both public databases. That code arrives as a card, too, and the card shows the address it will call. Here it is UniProt's public server. We allowed it once. This page lists the websites its code may reach. UniProt and the PDB are on it, so no extra network card appeared. The result opens in a built-in viewer. Drag and it turns, human hemoglobin entry 4HHB with four chains, two alpha and two beta. The identifiers it used P68871 and 4HHB.

It remembers and it counts

[3:57] The app also remembers after these runs it wrote notes of its own into a memory for this project where the file lives with the fit gave how the structure was fetched. Memory changes what happens next in a new session with the same question. Claude reused the earlier chart style without being asked and saved version three of the same figure. A second run is not a clean test. And it counts usage credits were off for the whole tour. The session meter read 7% when it began. It read 10% when the tour ended. No prompt ever asked for credits. Heavy work can go elsewhere. Settings compute lists SSH hosts, cloud providers, and model endpoints. We only looked no host was set up. This is where a cluster would be added.

Recap

[4:40] Let's recap with Claude. Claude science runs on your own computer with its sandbox on. Every command and every folder arrives as a card you answer first. Every result is an artifact with versions, a provenance record, and a reviewer. And skills, connectors, and a viewer reach real data while memory and usage show the cost.

Your turn

[4:58] Your turn. Paste this into Claude science. Read one small public data file from a folder I grant you read only. Fit a simple trend to it and save a figure as an artifact named trend.png. Then redraw it with a logarithmic x-axis under the same name and tell me what you are unsure of. Then check two things yourself. Did every code card show the code before it ran? And does the figure open with two versions you can step between? What is Claude science at Nick Bear Brown?