Coding with AI

I have written a couple of posts about cons of blind use of AI for writing code for Bioinformatics. Don't take me on the AI bashing side, I myself use Claude Code and Cursor. The AI coding assistants help me writing code faster for sure.

Here is how I extract the most out of them reliably.

1. 𝗦𝗲𝘁 𝘂𝗽 𝘆𝗼𝘂𝗿 𝗮𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁𝘀 𝗽𝗿𝗼𝗽𝗲𝗿𝗹𝘆 𝗳𝗼𝗿 𝗮 𝗽𝗿𝗼𝗷𝗲𝗰𝘁
Before starting on a project, spend time in defining a plan. Could be in form of claude.md in Claude Code or .cursorrules in case of Cursor. Mention all considerations, for ex: folder structure, test cases, overall intent, rules etc. Do these docs well with examples, intent and structure.

2. 𝗦𝘁𝗮𝗿𝘁 𝘄𝗶𝘁𝗵 𝗮 𝗰𝗹𝗲𝗮𝗿 𝘁𝗮𝘀𝗸, 𝗻𝗼𝘁 𝗮 𝘃𝗮𝗴𝘂𝗲 𝘄𝗶𝘀𝗵
Even for specific tasks, instead of "help with this bug", I describe the goal and constraints.
"Given this function, add input validation and clear error msg. Do not change the public API."
Task first prompts consistently produce better code.

3. 𝗙𝗲𝗲𝗱 𝗶𝘁 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝘀𝗹𝗶𝗰𝗲 𝗼𝗳 𝗰𝗼𝗻𝘁𝗲𝘅𝘁, 𝗻𝗼𝘁 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗿𝗲𝗽𝗼𝘀𝗶𝘁𝗼𝗿𝘆
I open the relevant files, select the function/module, and add a comment about what it should do and then let IDE take care of it. Assistants perform best when the context is small and focused, not with a dump of thousands of lines.

4. 𝗗𝗼𝗻'𝘁 𝗮𝗰𝗰𝗲𝗽𝘁 𝘁𝗵𝗲 𝘀𝘂𝗴𝗴𝗲𝘀𝘁𝗶𝗼𝗻𝘀 𝗯𝘆 𝗔𝗜 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁𝘀 𝗯𝗹𝗶𝗻𝗱𝗹𝘆.
Developers often accept AI suggestions with less scrutiny than they would from a human. Don't do that. Ask the assistants to explain the approach & reasoning, consider alternatives and weigh pros & cons. Treat the assistant like a smart junior developer. Always assume "this is a good draft, but I must verify it".

5. 𝗡𝗲𝘃𝗲𝗿 𝗰𝗼𝗺𝗺𝗶𝘁 𝗰𝗼𝗱𝗲 𝘆𝗼𝘂 𝗱𝗼 𝗻𝗼𝘁 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱
Big rule. If I cannot explain what the AI generated line by line, I do not ship it. I step through the logic, run it in a small sandbox, and only then let it near main.

6. 𝗞𝗲𝗲𝗽 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗮𝗻𝗱 𝗽𝗿𝗶𝘃𝗮𝗰𝘆 𝗶𝗻 𝘁𝗵𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝘆𝗼𝘂 𝘀𝗵𝗮𝗿𝗲
I avoid pasting secrets, API keys, real customer data or anything sensitive into prompts or shared traces. I use example data and clearly label it as such.

7. 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝘂𝗻𝗱𝗮𝗺𝗲𝗻𝘁𝗮𝗹𝘀 𝘀𝘁𝗶𝗹𝗹 𝗺𝗮𝘁𝘁𝗲𝗿 𝗲𝘃𝗲𝗻 𝗺𝗼𝗿𝗲
Plans, tests, good code practices, documentation and code review habits matter even more when a model is helping you write code. Be the driver behind the AI, don't let it make decisions on an autopilot. Be responsible for quality and maintainability.

Used this way, Claude Code, Cursor or any other IDE help you write the kind of code you are proud about, just with less grunt work and more focus on the interesting parts.