An unfamiliar equation, an error in reproduction code, or an idea you want to discuss can all be reasons to ask AI for help. Coding assistants can also read a project, edit code, and run programs. With your research material, they can help organize notes, analyze results, and make figures.
Coding and experiments
TRAE, WorkBuddy, Claude Code, ZCode, and Codex can help with project development. When reproducing a paper, you can provide the paper and code and ask for help understanding the implementation, setting up the environment, editing scripts, or diagnosing errors. A change you want to try can then be implemented in the existing project.
Once experiments are running, AI can help organize logs, compare results across configurations, and plot curves. Include the data locations, environment, and available compute when describing the task.
Research material and analysis
A discussion of a paper can draw on its text, equations, method diagrams, and code. With a collection of material, you can provide your research question and notes and ask AI to compare approaches, organize relevant experiments, and identify questions that remain unclear. The paper note workflow covers Zotero, Obsidian, and local AI tools.
Markdown works well for keeping an analysis in one document. Text, equations, code, tables, and images can sit together and remain easy to edit. An experiment analysis might include comparison conditions, plots, unusual examples, and your current explanation, with the relevant parts updated as new results arrive. Markdown's basic syntax covers these formats.
Figures and presentations
AI can help turn existing analyses and material into figures, webpages, or slides. Plots can come from CSV files and experiment logs; model architectures can become diagrams. Interactive pages or animations can show how something changes over time.
For a group meeting or talk, you can provide the audience, available time, and research material to develop speaker notes and slides, then refine the order, colors, and layout. Related resources are in writing, figures, and talks.
Discussing methods and ideas
AI can also be useful when you have an early idea or several possible explanations for a result. Relevant papers, previous attempts, and current questions give the discussion something concrete to work with: connections between methods, factors you may have missed, and other approaches worth trying.
Grill Me uses follow-up questions to explore your understanding. Tools are available in paper-tools-marketplace, with an example in paper reading and discussion. Prompts, procedures, and output formats you use often can be saved as a skill.
Keeping your own judgment
In a research discussion, AI may follow your idea and supply a motivation, method, and plausible explanation. A smooth conversation can make gaps in the reasoning less noticeable, or make its agreement feel like confirmation that you are right. You still need to consider whether its reasoning holds and fits the papers you have read and experiments you have done. Disagreements can also be discussed with teachers and other students.
Your own learning, experience, and judgment matter. Reading papers, understanding code, and doing experiments shape the questions you can ask and how you assess a suggestion. When a gap in your knowledge gets in the way of understanding the work, it is worth spending time on it. That is part of developing your own view of a problem and its possible solutions.
Two observations about AI discuss belief reinforcement in AI interactions and how students use and understand AI.
Further reading
Lilian Weng's agent article covers planning, memory, and tool use; Building effective agents discusses how workflows and agents are organized.