Strategies for accelerating scientific software development with artificial intelligence

Brad T. Aagaard

Submitted August 30, 2026, SCEC Contribution #15501, 2026 SCEC Annual Meeting Poster #212

Artificial intelligence (AI) tools, such as large language models, can accelerate software development by suggesting designs and writing code. However, AI tools often make unstated assumptions and can generate confident but false information (“hallucinations”). This can lead to responses that are incomplete, wrong, or not what is needed. Many online guides offer tips, but few are tailored to scientists, and general claims about effectiveness are hard to verify independently. Drawing on practical experience, I present some general strategies for using AI tools to accelerate scientific software development, which could be useful in other areas as well.
1. Write prompts that provide clear, well-defined tasks: ask AI to state its assumptions (or ask questions if unsure); favor simple solutions over unneeded flexibility; limit changes to what is actually needed.
2. Build reusable "skills" (for example, a saved set of guidelines for reviewing code) to write well-defined tasks faster.
3. Refine a design to create a detailed specification before asking AI to write any code.
4. Ask the same question to multiple AI tools and combine the best parts of each answer.
5. Examine cases in which AI tools produce incorrect results or not what is desired to find weaknesses in your software's design.

I have found these strategies to be useful in
• Brainstorming simpler, more flexible designs for code components and libraries, including ones that use newer language features;
• Reviewing code for memory, security, maintainability, and performance issues;
• Writing test suites for many components from just a few examples;
• Writing consistent documentation for application programming interfaces;
• Finding documentation gaps for both new and experienced users; and
• Building the skeleton of a graphical user interface from a design specification and an existing application built using the same tools.

These strategies can also be applied to other tasks suitable for AI agents, such as translating code from one programming language to another and creating configuration files for tools that build and install software. In these tasks, I have found current AI tools to accelerate software development by approximately fourfold, which yields more time for designing rigorous tests.

Key Words
artificial intelligence, scientific software

Citation
Aagaard, B. T. (2026, 08). Strategies for accelerating scientific software development with artificial intelligence. Poster Presentation at 2026 SCEC Annual Meeting.

Related Projects & Working Groups
Research Computing (RC)