5 Tips for Using AI in research:

Coding assistants have recently taken the world by storm. Simply put, there are now extremely capable code writers which can work on large complex problems & code bases. The frontier of tools and techniques moves extremely fast. And in this blog we will explore the frontier!

But first things first, before we can talk in the future about advanced tools, let's start with the essentials first.

This post is to get you started to use coding assistants to implement your research projects in a sustainable way.

Let's dive right into it!

Gemini's tips (for fun):

Before going to the tips I collected in the last month, here are some tips generated by Gemini 3 Pro (low):

  1. Explain legacy code: Paste that cryptic Perl or R script from 2012 into the AI to get a line-by-line explanation of what it actually does.
  2. Generate plotting boilerplate: Describe the visualization you need (e.g., "Volcano plot with significant genes labeled in red") to get a working Matplotlib or ggplot2 starting point.
  3. Debug pipeline logs: Paste error output from Snakemake or Nextflow to identify missing dependencies or syntax errors faster.
  4. Format conversion: Quickly generate one-off scripts to convert between incompatible formats (like FASTA to CSV) without writing regex yourself.
  5. Summarize literature: Paste abstracts or full papers to extract key findings, methodologies, or parameter values for your own models.

My tips:

OK, Gemini clearly was biased in the tips it generated, I would say from my past projects. Here is the list I came up with:

1. The filesystem is your friend, and CLAUDE.md is the entry point

As AI models have a limited context window, it is a good strategy to store relevant information in files, which the AI agent can open (or ignore) as needed. I recommend creating a CLAUDE.md (or similar) file as the main entry point for persistent information and structure. Keep it simple and short initially, revisiting it when the scope changes.

2. Collaborative planning

Plan first. Think of the AI as a collaborator. It is an extremely versatile programmer, capable of high standards. Plan with an open mindset, explain your real goals, and work together with the agent, not just "above" it.

A specific feature of Claude Code is the "Plan Mode" which helps you to structure your thoughts before writing any code. You can enter it by typing:

/plan

This is especially useful for complex tasks where you want to agree on a strategy first.

3. Use git, your safety net!

Don't rush ahead. To keep a history of all changes, use GIT.

  • Initialize
  • Add changes
  • Commit! Do this after each step.

In the beginning, manual commits are recommended to make sure you are mindful about the changes added in each step. References: VS Code Git Version Control

4. Avoid the stupid zone

Keep sessions dedicated to their own topic. If one session is overly long, the agent will start to over-emphasize small details or lose context. Start fresh when shifting focus.

5. Be mindful and slow

Instead of driving blindly to just do more and faster, use this opportunity to do more meaningful work. Use the time saved to deeply focus on the right questions, to be attentive and curious.

This is just the starting point. My research in the next months is aimed to catch up with this rapidly evolving field. Stay tuned for future posts exploring this space together with us.

Finally, a word of caution is in order: As a community of researchers, we should use this new resource wisely. Instead of driving blindly in just doing more, publishing more papers and writing programs faster, I think this is now the opportunity to do use this new time to do more meaningful work, in times of attention decay, imagine how much value we can add if we use this time to deeply focus on the right questions, to be attentive and curious.

Note: I use Claude Code CLI, which I personally prefer for its command-line efficiency and deep integration with my development workflow. To use Claude Code, you will need a paid subscription to Anthropic's services or sufficient API credits. Check their pricing page for the latest details on subscription models. Official Documentation: Claude Code Docs