Over the last 6-8 months, I've been using GitHub Copilot regularly and have narrowed it down to two ways I rely on it. Not for autocomplete. But as a thinking and execution partner. 1️⃣ Understanding existing code When I need to understand unfamiliar or complex code, I use Ask mode. What I usually do: - Share the code snippet and any references I already know are relevant. - Before asking for an explanation, I ask Copilot what additional references or context it needs to explain the code better. - Once all required context is attached, I ask it to explain the code step by step. This approach reduces guesswork and leads to much more accurate explanations, especially in large codebases. 2️⃣ From requirements to code This is where Copilot helps me the most. Step 1: Clarifying the problem I paste the requirements into Ask mode and start a discussion: - What approach should we take? - What assumptions are we making? - What are the edge cases? There's a lot of back-and-forth here. Sometimes I explain why something won't work. Sometimes Copilot points out gaps in my reasoning. Step 2: Confirming understanding Before any code is written, I ask Copilot to: - Re-list my requirements (to confirm it understood them correctly) - Explain the approach it plans to take - List the files / classes / functions that will be created or modified - Break everything down into numbered tasks Only after this do I move forward. Step 3: Writing the code I then switch to Agent mode and use premium models (Claude Sonnet 4 earlier, Sonnet 4.5 lately) to implement the tasks. Step 4: Review and iteration I first read and validate the generated code myself. Based on that validation, I decide how to proceed: - If the change is small, I make it manually - If it's a logic issue, I point it out and ask Copilot to fix it - If it's a bigger issue, I switch back to Ask mode to discuss the correction or approach, and then either apply the fix myself or switch back to Agent mode to update the code This loop continues until I'm satisfied. What this taught me Copilot works best when you stay in control and treat it like a collaborator, not a replacement. #GitHubCopilot #VSCode #DeveloperProductivity
Adarsh Nandigama
2mo
In prompt engineering they call it “Flipped Interaction” After describing the desired output, We ask LLM to ask the questions regarding the requirements to compelte the given task without assuming It decreases re-asking to tune the output and get the concise output at once. Nice post Bingi Nagesh
Very good practical insights.
Bingi Nagesh for more structured approach, feel free to use official spec kit guidance. https://github.com/github/spec-kit