With the rapid improvement of AI models, vibe coding or context coding has become increasingly popular. In simple terms, it’s writing code with the help of AI.
After using GitHub Copilot extensively, I’ve developed a workflow that consistently delivers better results. I’d like to share it here.
These experiences are based on using GitHub Copilot in Visual Studio Code, but I believe most of the principles apply to other AI coding tools as well.
1. Start with a New Branch and a New Chat
Whenever you need to implement new requirements or fix bugs, create a new Git branch first. Then start a new Copilot chat dedicated to that branch.
If the task is complex, pin the chat and treat it as the working context for that branch.
In reality, we often work on multiple tasks within the same codebase. You may switch between branches throughout the day, even if a Copilot conversation isn’t finished. Just make sure the correct chat stays associated with the correct branch.
Keeping branch context and conversation context aligned helps prevent confusion and improves AI-generated suggestions.
2. Always Begin with “Plan” Mode
Before asking Copilot to write code, switch to Plan Mode.
Describe the requirement, bug, or modification you want to make. Sometimes Copilot will ask follow-up questions. Answer them as accurately as possible.
If you’re unsure about an answer, don’t guess. Pause, investigate the issue, gather the missing information, and then continue.
The quality of the output depends heavily on the quality of the context you provide.
3. Review the Plan Carefully
Once Copilot finishes planning, review the proposed approach thoroughly.
If anything is unclear, ask questions and request clarification. Don’t switch to Agent Mode until you’re confident the plan is correct.
A typical workflow may look like this:
Plan → Review → Ask Questions → Refined Plan → Review → More Questions → Final Plan
This process can take several iterations, and that’s perfectly normal.
In fact, when the plan is long and detailed, it’s often a good time to take a short break. Grab a drink, walk around, or switch to another small task for a few minutes.
When we’re eager to solve a problem quickly, it’s easy to overlook critical details. Slowing down often leads to better AI-assisted coding results.
One habit that has helped me a lot is reviewing the entire plan before proceeding. If Copilot only shows part of the plan, ask it to display the complete version.
4. Switch to Agent Mode
Congratulations. The most demanding part is usually done.
Once you’re confident that the plan is solid, switch to Agent Mode and let Copilot implement the changes.
At this point, Copilot has a much higher chance of generating useful code because both you and the AI are aligned on the expected outcome.
5. Review, Test, and Iterate
After the implementation is complete, review the code and run your tests.
If the result already meets around 60% of your expectations, commit the work to Git first.
Then continue refining it with Copilot. Explain what isn’t working, what needs improvement, or what edge cases were missed.
At this stage, either Plan Mode or Agent Mode can work well because the overall direction is already correct. For smaller adjustments, Agent Mode is usually sufficient.
Summary
I know there are many advanced techniques for using Copilot, such as Skills, specialized markdown templates.
However, the process above is enough to solve most of my daily coding challenges.
The key lesson I’ve learned is simple:
You must understand the problem and its background context before asking Copilot to implement a solution.
Doing the right thing is always more important than doing things right quickly.
That’s why I believe the Planning Phase is the most critical part of AI-assisted coding. A good plan dramatically increases the quality of the final result.