10 ways to use Claude for scientific work

Recurring tasks you set once; the same research prompt then runs on a schedule and reports what changed. Ours is a weekly trigger that scans for every company with a drug targeting the folate receptor (FRα/FOLR1) in preclinical or clinical development, compares against last week’s answer, and reports the delta:

Every Monday morning: search ClinicalTrials.gov, company press releases, and recent conference abstracts for drug programs targeting folate receptor alpha (FOLR1), preclinical or clinical. Compare with last week’s list. Report: new programs, phase transitions, discontinuations, and notable data readouts, each with a source link.

Competitive intelligence stops being a task someone forgets and becomes a feed. The same pattern works for new papers in your field, preprints citing your target, or regulatory decisions in your indication.

Connectors (built on MCP, the open Model Context Protocol) plug Claude directly into external tools and databases. The connector directory has a life-sciences section: PubMed, ClinicalTrials.gov, bioRxiv and medRxiv, ChEMBL, Open Targets, and more.

General web search skims; connectors query. With PubMed connected, Claude retrieves actual article records; with the Clinical Trials connector it filters studies by phase, sponsor, and status; with ChEMBL it pulls bioactivity data. A literature dive that used to be twenty browser tabs becomes one conversation with sources attached.

Skills are packaged instructions Claude loads on demand; in the app you invoke them with a slash command. Compare the same question asked two ways. Plain chat: “which tools would be best to design a new compound against colitis” gets you a competent wall of text. With skills, the prompt becomes:

With the skills loaded, Claude pulls in the relevant connectors and then interviews you: what modality, what stage, what data you already have, what decision you are actually trying to make. The output is a starting point for drug design, not an essay to digest. The difference between an answer and a workflow is the questions asked before the answer.

A Project is a workspace holding your notes, papers, and instructions; every chat inside it starts with that context loaded, and the project accumulates memory across chats.

Our “Investigating agents against IBD” project holds the key review PDFs, internal notes, and instructions on how we want analyses framed. Chats like “Promising compounds in clinical trials” and “Approved drugs for ulcerative colitis” share that context instead of restating it, and connectors can be approved once for everything inside the project. Context is the tax you pay on every new chat; projects pay it once.

With Claude in Chrome, Claude operates a real browser: clicking, filtering, and navigating web tools rather than just fetching pages. Ask it to open NCBI and actively apply the same filters you would use by hand (organism, molecule type, date range), then walk the results with you. An enormous amount of scientific tooling lives behind web UIs with no API; browser control makes those tools scriptable in plain English.

In the Claude desktop app, Chat is the conversational mode you know. Cowork gives Claude a working session: access to your files and the ability to run longer, multi-step autonomous work that ends in deliverables. Chat is for questions (“what is known about TL1A in IBD?”); Cowork is for jobs (“go through these 40 PDFs, extract every reported EC50 into one table, and flag inconsistent units”). Most disappointment with AI assistants is a mode mismatch: asking a chat window to do a job, or spinning up a working session for a one-line question.

Claude produces complete, editable slide decks: journal-club summaries, project updates, conference backups. Claude Design generates visual layouts you refine element by element and exports to PDF, PowerPoint, or HTML; Cowork with the pptx skill produces PowerPoint files directly when you want the file, not the canvas. Paste the paper or point at the project, describe the audience (“15 minutes, lab meeting, emphasize methods”), and iterate on the deck instead of building it. The first 80 percent of most scientific decks is layout labor, not thinking.

Combine an ELN connector with literature access and a protocol becomes checkable before it runs. Benchling ships an official Claude connector for exactly this kind of work: pull the protocol from your ELN, then ask Claude to compare it against published methods: missing controls, unusual concentrations, steps that published papers do differently, reagent substitutions reported to fail. Ten minutes of review before a two-week experiment. Most protocol failures are known failures, published somewhere you did not look.

The same Benchling connection works in the other direction: Claude reads the actual experimental records (entries, results, sample metadata) and drafts the write-up. “Summarize the last three weeks of the conjugation-optimization entries into a results section with a table of yields” is a realistic prompt when the ELN is connected; the draft cites the entries it drew from, and the scientist edits science instead of typing structure. Writing up is where finished science goes to wait, and shortening that queue compounds.

Claude Code is Anthropic’s coding agent: it writes, runs, and iterates on real software in your repositories, from the terminal, desktop, or web. The repetitive workflows every lab tolerates (renaming and merging instrument exports, converting between formats, a small dashboard for assay QC, batch-processing sequencing outputs) are each an afternoon of Claude Code, not a quarter of an engineer’s year. This is the one item on this list that is qualitatively hard in a chat window: chat can write a script, but Claude Code runs it, sees the error, fixes it, and ships a working tool with tests.

The catch, and the reason this list ends here: code alone is not the goal. A generated script is only as good as the scientific judgment wired into it, which is why the strongest setup pairs the coding agent with preset scientific skills: Anthropic’s bio-research plugin, for example, bundles ten scientific database connectors and six analysis skills for sequencing, drug discovery, and single-cell work. And when the job is not “build me a tool” but “answer this research question with validated analysis,” that is a different product category altogether.

The ten workflows above make a general-purpose assistant genuinely useful in a lab, and the direction of travel is clear: even Anthropic now ships a dedicated research workbench (Claude Science, in beta) because science needs more than a chat box. But every workflow on this list shares one limit: you are still the orchestrator, the validator, and the one who knows whether the analysis was done right.

That second layer is what Hydra is: give it a research direction and it plans the analysis, runs it across ~100 scientific databases and 200+ codified bioinformatics skills, and validates every result before you see it. If the ten workflows above are your on-ramp, try Hydra when the question matters more than the tooling, or see the case studies for what that looks like in practice.