AI Workflow

I have been using to do my professional work since 2 years when Github Copilot was launched and we got an official subscription from work. Since they I have used Claude, ChatGPT and more recently Cursor AI.

However, these days I am getting mixed feelings about it. I believe getting AI to work in Agent Mode will basically cause brain atrophy and you will loose your ability to reason and program well. So solve this, I would use AI more as a pair programmer or research assistant and not as a code generator. And even in the case, it generates the code it has to to pass all benchmarks and code standards. It is very easy to get overwhelmed by AI and the amount of code it generates.

So from now on, this will be my workflow:

• Take the problem statement and think hard about it. Come up with a plan.
• Ask the same problem to the AI and use their thinking mode to come up with a plan. See if the plan is similar to mine and if mine can use some ideas from it.
• Once the plan is complete, start writing the code with a fast feedback loop and preferably in a test driven way to ensure the code written is correct.
• Once the code is written, ask the AI to review the code and suggest improvements. Again, no agentic mode here and I want to write all the code committed from my name. AI can be used to generate documentation but has to be critically reviewed.
• Finally, make a MR/PR and then get the review done and there would be cycle of QAs to check my work.
• I believe I can use the AI to its strengths and at the same time now weaken my abilities by over relying on it.

That’s it folks! Maybe you get smarter and prosper.