🧠 Letting AI Drive: A Month of Building with GitHub Copilot

“When the going gets weird, the weird turn pro.”

— Hunter S. Thompson

What happens when you stop writing code and start managing your AI developer instead?

⚡ TL;DR;

I gave GitHub Copilot full control of a built-from-scratch NodeJS project for a month.

ChatGPT handled planning, strategy, UX input, and story writing.

I acted as manager, director, and chaos wrangler.

The result:

A working prototype, repeatable prompt patterns, and serious lessons on using AI as your dev partner.

👋 Setup

Built as a personal experiment on a work machine (private, for now).

Toolchain:

• VS Code Insiders, started in Codespaces for safety
• Full access granted to terminal, extensions, file tree, everything
• Linting + formatting: ESLint, Prettier, Remark, Commitlint (strictly enforced)
• Testing: Vitest with 90%+ coverage enforcement
• Copilot handled: Code, tests, docs, commit messages, pull requests, story implementation
• ChatGPT handled: Story writing, backlog creation, research, branding
• Workflow: Start with instructions (repo-level + personal) → evolve into structured prompts → iterate and commit

Most //human input lines were documentation. Actual code edits? Maybe two — one for manual testing and one was copy/pasted.

🧪 How I Actually Prompt (GitHub Copilot)

🧠 Copilot Agent Settings

Note: By default, Ask mode doesn't save historical context

Depending on what I’m doing, I pick a specific model under Agent mode:

• 🛠️ Feature implementation, heavy refactor, full A→Q problem-solving

  • Claude Sonnet 4 or Claude Sonnet 3.7

• 🧪 Testing, docs, explanation, general idea bouncing

  • GPT-4.1 or GPT-4o

• 🔍 Somewhere in between

  • Gemini 2.5 Pro (when it actually works)

• ⚙️ Light scripting, config setup, follow-along patterns

  • o4-mini

🧷 Most-Used Contexts

• #terminalLastCommand
• #selection
• #file

🧠 My Go-To Prompts

When I’m Stuck

#codebase I want to build a <thing> for <goal>.  Given current best practices in <language>, analyze the top 3 ways to implement it. List 2 pros/cons for each. Rate 1–5 for complexity, ease of use, maintainability. Rank them and explain why. 

Even when the simplest ones didn’t work right away, they always got the ball rolling. If needed, I clarified or reset with /clear and tried again.

Doesn't Compile

/fix #the-file-copilot-prob-broke-anyway.js 

Failing Tests

/fix #terminalLastCommand 

Lint errors

/fix #problems 

☠️ The Copilot Hell Loop Is Real

It happens: Copilot spirals into broken logic and keeps rewriting working code.

Here’s what works:

Pause → Redirect → Resume

(But don’t wait too long — or your fix becomes git reset --hard HEAD.)

📚 What I Learned

✅ Models Matter

Pick the right one — it'll save you hours of rework.

✅ Specific Prompts Win

Copilot thrives on intent and clarity, not vague instructions.

✅ Back-and-Forth Works

#selection I was thinking of doing X here — would you recommend that?

Copilot actually responds meaningfully.

✅ Commit While It Works

Don’t wait for perfect. Save progress, iterate fast.

✅ No Plan? No Problem (Sometimes)

Just stay curious and let your Copilot riff with you.

❌ #codebase for Direct Implementation Doesn't Work (for Me)

I know it's a headline feature, but asking Copilot to /implement-feature JIRA-123 #codebase has never worked well for me.

What does work? Asking targeted questions, especially during active review or when you’re trying to break down next steps.

⚠️ Long-Running Chat Histories Hit a Wall

Even if you think building a giant rolling context will help, Copilot's memory eventually falls off.

What worked better:

Story → review → refine → commit → /clear

Repeat that and results stayed sharp.

🎤 Try It: VS Code Speech for Conversational Prompts

VS Code Speech gives you a “hey code!” style interaction with Copilot.

You can speak out goals, requests, or reviews and let it translate to code suggestions.

⚠️ Just watch your token usage — premium-style interactions can balloon into fast rate limits if you're not careful.

✅ ChatGPT = Worth the subscription

Docs, strategies, structure, creative writing, even “Hey, does this make sense?” moments.

We can't always live in code (plus, who can resist playing with pictures and color?)

Yes, it helped with this post too!

✨ Let’s Talk

Questions? Want to swap prompts, patterns, or war stories? Hit me up — just give me a few weeks to recover.

I’m still prepping for the internal demo and approvals.

This was Part 1 of a 2-part breakdown.

Next up: The build, the metrics, and what 600,000+ lines of AI-generated code actually look like in the wild. Letting AI Drive: From Hell Loops to Happy Commits.

🛡️ RAI Disclaimer

Everything I share here is my own perspective—created with the help of AI tools (GitHub Copilot, ChatGPT, and their friends), but always with a human in the loop. I do my best to catch accidental bias and fact-check, but if you ever spot something odd, let me know! AI isn’t perfect, and neither am I.

TL; DR: AI helped, but you can blame me for the chaos! 🫠