Claude Code for Researchers: Build a Research Agent Without Writing Code

In short: Claude Code is an AI tool that works inside a folder on your computer: it can read your files, follow written instructions and carry out multi-step tasks like sorting a folder of papers into a table. You don't need to code. You need a clear folder, a written instruction file (CLAUDE.md), a small first task and the habit of reviewing what it did.

Why researchers hit a wall with chat

You have 60 PDFs, a spreadsheet of results and a protocol document. In a chat window you paste a few pages at a time, lose context, re-explain your project and copy answers back by hand. The model is capable, but the setup makes you the courier.

An agent changes the setup. Instead of you carrying text to the model, the model works where your files are: it opens them, reads them, writes new files and tells you what it changed. Claude Code, from Anthropic, is one such tool. It started as a coding assistant, but nothing about the pattern is limited to code.

What a "research agent" actually is

In plain terms, an agent is a model plus four things: instructions it reads at the start, files it can see, tools it may use (read, write, search, run a command) and permission rules that say what it must ask about first. A research agent is just that setup pointed at research work.

What it can do with a project folder

• Read a folder of PDFs and produce a structured table: design, sample, method, main result, limitations, with the supporting quote for each.
• Draft a methods paragraph from your protocol notes, keeping to what the notes say and flagging what is missing.
• Check a spreadsheet for inconsistencies: mixed units, duplicated IDs, impossible values. It reports them, and you decide what to do.
• Rename and organize files by a convention you define, and keep a log of what it moved.
• Turn lab-meeting notes into an action list that cites the file each item came from.

What it should not do on its own: choose a statistical test for your data, decide what a result means, or touch raw data you have no copy of. Its output is a draft for you to review.

Set it up in five steps

1. Start with a clean folder

Create one project folder with a copy of what the agent needs, not your whole drive. For example: a papers folder, a protocols folder, a data-copy folder and an empty output folder. The agent works inside this folder, so the folder is your first safety boundary.

2. Install it from the official instructions

Installation steps and plan requirements change, so follow Anthropic's current documentation rather than a blog post. Expect to need a paid Claude plan or API access.

3. Write a CLAUDE.md file

CLAUDE.md is a plain text file at the top of your folder that Claude Code reads at the start of each session. It is where you explain your project once, instead of every time. It is the single most useful thing you can write.

# Project
I study [topic] in [system/organism/patient population]. This folder holds [papers / protocols / data copies].

# How to work
- Use only the files in this folder. If something isn't in them, say so; don't fill the gap from memory.
- For every extracted fact, give the file name and a short verbatim quote.
- Never modify or delete files in papers/ or data-copy/. Write new files only in output/.
- Write for a [postdoc / clinician / R&D manager]. Plain language, no hype.

# Output conventions
- Tables as CSV or Markdown. File names: YYYY-MM-DD_short-description.
- End every task with a list of what you did, what you were unsure about and what I should check.

4. Give it a small, read-only first task

Read the PDFs in papers/. Create output/summary-table.csv with one row per paper and these columns: file name, study design, sample or model, key method, main result with numbers, stated limitations, and a verbatim quote supporting the main result.
Don't change any existing file. When you finish, list any paper you could not read or were unsure about.

5. Review before you trust

Open the output and check a sample against the source, the same way you would check a student's first extraction. Most problems come from an ambiguous instruction, so when something is wrong, fix CLAUDE.md rather than just correcting the one output. The agent then improves for the next run. Putting the folder under version control, for example with Git and GitHub, gives you a record of every change it makes and an easy way to undo one.

Safety and honesty about limits

• Confidential data. Do not point an agent at patient data, unpublished results or proprietary sequences unless your organization has approved the tool and the data handling. When in doubt, ask your IT or compliance team first.
• Permissions. Claude Code asks before it edits files or runs commands. Read those prompts; don't approve everything on autopilot.
• Errors are normal. Agents misread tables, miss a paper or state something too strongly. This is why every output carries quotes and a check list.
• Reproducibility. Treat the agent's output as a draft or a working file. Record which version of the instructions produced it.
• Cost and speed. Large folders use more usage. Start small.

Where to go from here

Once the first task works, add one skill at a time: a second task with its own instructions, a reusable checklist, a connection to your reference manager's export. Keep each step small and checkable. For the prompts that sit inside those instructions, see How to Write Good AI Prompts for Scientific Work, and for the extraction logic in the example above, AI Literature Review Without Fake Citations.