What One Week with GitHub Copilot Taught Me

Welcome to issue #71 of the iOS Coffee Break Newsletter 📬.

Until recently, I was still skeptical about using AI for coding!

I think part of that skepticism came from not fully trusting the quality of the output. Another part came from not wanting to become dependent on suggestions I did not fully understand. And, if I am being honest, I also felt these tools often looked much better in demos than they did in real project work.

That started to change this past week.

For a full week, I used GitHub Copilot much more intentionally in my day-to-day work. Not just for quick autocomplete suggestions, but as part of the actual development flow: asking it to help me reason about changes, break down tasks and move faster through implementation.

By the end of the week, my perspective was different.

This edition is a reflection on that experience: what worked, what did not, and what helped me move from resistance to a more practical mindset. I increasingly feel that this is something we should learn to work with instead of spending too much energy fighting against it.

These thoughts come from my experience using the GitHub Copilot Business plan. Do not take them as granted, they are just honest notes from my own daily workflow.

My setup during this week

At first, I wanted to use Copilot straight from Xcode through CopilotForXcode.

I gave it a fair try, but it still does not feel reliable enough for my day-to-day work. What ended up working better for me was this:

• Xcode for my regular development work
• Visual Studio Code when I want to use the agent in a more intentional way

It is not the ideal setup, but it has been much more dependable.

What helped the most

The one thing that improved the experience the most was this: start a new conversation for each new feature.

I noticed that when a conversation became too long, the agent gradually lost focus. In some cases, that also led to hallucinations.

Keeping each conversation tied to a specific goal gave me much better and more predictable results.

Where I went wrong more than once

A mistake I made several times was trying to get too much done in a single prompt.

When I asked for a large change all at once, the agent often struggled to stay focused. I would see it:

• touch files I did not mean to change
• make edits unrelated to the original task
• move away from the outcome I actually wanted

Now, whenever I notice that happening, I pause and break the task into smaller parts.

What works better for me is:

• start by asking for a plan
• move through the work step by step
• review each step before going further

This made the workflow much safer and a lot less noisy.

MCP support ended up being really helpful

One other thing I explored this week was MCP support.

I set up the Atlassian MCP Server, and it worked surprisingly well for giving the agent richer context about my tasks.

Having Jira details available in the conversation made it easier for the agent to stay aligned with what I was actually trying to do. That mattered more than I expected, especially when the work involved product requirements, task history, or details that were not obvious from the code alone.

In practice, this was one of the improvements I found most valuable!

The biggest takeaway for me

If there is one thing this week made clear to me, it is this:

Make sure you understand the fix before accepting it.

If I cannot explain why a solution works, then I am not in a good position to validate it. And if I cannot validate it, I should not be relying on it.

Copilot can speed up the process, but it does not remove our responsibility to make sure the code is correct.

Final thoughts

After spending a week with GitHub Copilot in my iOS workflow, my impression is positive, though definitely more realistic now.

It can be a very useful tool when the problem is clearly defined and the context is solid. But once the scope gets too wide and validation becomes weaker, things can go off track quite quickly.

The biggest change for me is not that I now trust the tool blindly. It is that I now understand better where it fits in my workflow, and where I need to slow down and think for myself.

I plan to keep using it, just much more intentionally than I did on day one.

Have a great week ahead 🤎