Recording My Experiments in Spec-Driven Development

#programming #ai #jira #contextdrivendevelopment

For years, I’ve been chasing one idea — keeping context close to the work.

Whether it’s TODO comments, markdown notes, or tasks in the repo, my goal has always been to reduce the mental tax of switching tools just to remember what I was doing.

That’s the foundation of what I call context-driven development — a workflow where the “why” behind each change lives right next to the “how.”

Now I’m taking that same idea into the world of AI-assisted coding.

Recording My Experiments

Lately, I’ve been experimenting with spec-driven development — trying to guide AI tools like Copilot to produce code that matches what I actually want, not just what it infers from nearby syntax.

Progress has been slow.

There’s just too much going on in my life right now to hold the full context of what’s working and what isn’t. So I’ve decided to start recording my experiments and sharing them here.

Embedding Context Where It Belongs

In my latest experiment, I used context embedded in the Jira issue to guide Copilot.

Instead of writing long, open-ended prompts, I pasted that context into the Copilot chat panel and referenced specific files and lines.

That small change made a big difference. It gave Copilot something closer to what a teammate would see — the why behind the change, not just the where.

Moving Toward Structured Specs

While researching how ChatGPT interprets structured text, I landed on a markdown format that should be easier for Copilot to understand and follow.

Here’s the format I’ll be using in my next experiment:

- [ ] <task title>  
  - files: `<path:line>`, `<path:line>`  
  - goal: <what the change achieves>  
  - constraints: <rules to avoid unwanted changes>  
  - action: apply patch and show diffs

My hope is that this will make the generated code more predictable — and closer to what I’d write myself.

What’s Next

In the next post, I’ll use this format to see how well Copilot follows the spec when it’s embedded directly in the codebase.
I’ll record the whole thing so you can see exactly how it performs.

Stay tuned — and if you’ve tried something similar with Copilot or ChatGPT, I’d love to hear how you structure your specs.

Watch the short intro video here:
👉 https://youtu.be/Op9iQ2LX2DI