3 min read

💥 I Let Copilot Write My Code for a Week. Here’s What Broke.

💥 I Let Copilot Write My Code for a Week. Here’s What Broke.
Photo by Jay Heike / Unsplash

Spoiler alert: It wasn’t just my build pipeline.

🤖 Day 0: The Glorious (Delusional) Plan

Like every ambitious developer who’s tired of typing map(item => item.name) for the 9,000th time, I thought:

“What if I just let GitHub Copilot write everything for me this week?”

It’s 2025. We’ve got ChatGPT, Cursor, WindSurf, even AI agents who know what coffee you like before you do. Why not outsource my job to a cheerful autocomplete?

I told my team I was experimenting with AI-assisted workflows. I told myself I was being innovative. What I didn’t tell anyone was: I had absolutely no idea what I was doing.


⚙️ My Setup (aka “I Built the Perfect Dev Trap”)

  • IDE: VS Code (because Copilot lives here rent-free)
  • Copilot: Fully activated, default config
  • Cursor: Installed and lurking in the background for backup
  • ChatGPT Plus: On tab #3 like a pocket philosopher
  • WindSurf: For semantic code search and pretending I have a system

🚧 Day 1: The Typing Stops, the Guessing Begins

What happened: I stopped typing like a real developer and started pressing Tab like a caffeine-deprived monkey on autopilot.

What broke: My brain.

Copilot tried to complete everything. Even things I didn’t mean to write.

Example:

// Wanted: check if user is admin
if (user.role === 'admin') { ... }

// Got:
if (user.role === 'admin' || 'superadmin' || 'godmode') { ... }

😬 Guess Copilot thought I was building Skynet.


🤹‍♀️ Day 2: The AI Improv Comedy Show

What happened: Copilot started guessing my intent... poorly.

Me: “Let’s write a function to validate email input.”

Copilot: “Here’s a 12-line regex that looks like it came from the Matrix.”

Me: “Uh… what’s this line with 🦄?”

Even ChatGPT, when I copied and pasted the mess in, said:

“I’m not sure where this came from. Try again?”

🧪 Day 3: Unit Tests Written By a Goldfish

Thought I was winning when Copilot auto-generated unit tests.

Turns out, they were:

  • Testing the wrong function names
  • Missing important edge cases
  • Passing even when the actual logic was failing

Prompt that helped:

“Write a Jest unit test for a function that takes user input and returns true if it matches a 10-digit number. Include tests for empty strings, null, and undefined.”

Thank you, Cursor. You saved me today.


🧼 Day 4: The Great Refactor (aka The Great Undo)

I let Copilot refactor a few components. It removed “redundant” code.

One of them was a logging statement that caught a payment failure.

Oops.

You don’t know terror until you push to staging and Stripe logs go silent.

Lesson: AI doesn’t know which "unused variables" you care about. Humans sometimes do. Sometimes.


🧠 Day 5: Meetings > Coding?

Ironically, the one place AI saved me this week: not code.

I asked ChatGPT to:

  • Summarize our last engineering sync
  • Generate a Slack message for a delayed migration
  • Rewrite my retro notes to make them sound less angry

Prompt that worked best:

“Here’s a meeting transcript. Can you summarize key points, action items, and make me sound 15% smarter?”

I felt like the office Yoda. Calm. Wise. Composed. 100% fake.


🔥 Day 6: AI-Generated Burnout

I realized I was constantly second-guessing Copilot.

“Should I trust this function?”
“Did it import the right library?”
“Why is it suggesting SQL when I’m writing CSS?”

Turns out, outsourcing your code still requires... reviewing all the code.

Also turns out, pressing Tab 800 times a day is weirdly exhausting.


🌅 Day 7: Existential Dread, but With Syntax Highlighting

Final task of the week: fix a bug Copilot introduced on Day 2.

Spent 3 hours debugging a NaN issue only to find:

const quantity = parseInt(quantity) || 0;

✨ Infinite recursion, brought to you by: trusting autocomplete too much.

I let Cursor explain it to me. I let ChatGPT rewrite it.
I let WindSurf tell me which PR it came from.

I eventually just fixed it manually. Like a caveman. A wise, humbled caveman.


🛠️ TL;DR – What Broke (and What Didn’t)

What Worked 🚀 What Broke 💥
Using Copilot for boilerplate Trusting it for business logic
ChatGPT for writing/refining Copilot’s test generation (mid at best)
Cursor for full-project refactor Refactoring without context
Prompts with constraints Blindly tabbing through suggestions
AI for communication & retros AI as a replacement for thinking

🤖 Final Verdict: Use AI Like a Senior Would

Copilot is a tool — not a replacement for engineering thinking.

Use it to:

  • Save keystrokes
  • Reduce boilerplate
  • Kick off ideas

Don’t use it to:

  • Replace your judgment
  • Skip code reviews
  • Avoid understanding your stack

If you treat Copilot like a pair programmer with selective amnesia and a flair for overconfidence, you’ll be just fine.

If you expect it to own your output... let’s just say your Jira board’s about to get a lot more colorful.