AI Wrote Half My Week — Here's What the Numbers Actually Looked Like (Week 8 Roundup)

#ai #productivity #career #engineering

AI Wrote Half My Week — Here's What the Numbers Actually Looked Like (Week 8 Roundup)

This week had a clear through-line: using AI not as a novelty, but as a multiplier on two of the most time-consuming communication tasks in engineering — refactoring legacy code and writing pull request descriptions.

Here's what actually happened, with real numbers.

The Refactoring Session

On a recent project, I had a service with a module that hadn't been meaningfully touched in about three years. Nested conditionals, no tests, naming that made sense to someone in 2021 but nobody since. I ran it through a structured AI refactoring workflow — the kind I wrote about this week — and tracked the time.

Manual refactoring estimate (based on similar past work): ~4 hours.

Actual time with AI-assisted prompts, including review and sanity-checking the output: 1 hour 20 minutes.

That's not magic. That's systematic prompting: ask for a plain-language summary first, identify the hotspots, then refactor incrementally rather than all at once. The AI didn't understand the business logic — I did. It handled the mechanical transformation; I handled the judgment calls.

The PR Description Problem

The second theme this week was pull request descriptions. On the same project, I counted: over the last sprint, roughly 60% of PR review cycles included at least one round-trip comment that existed purely because the description was unclear — missing context, no "why this change," no callouts for reviewers.

I switched to generating first-draft PR descriptions with AI (prompt: diff in, structured description out), then editing for accuracy. Review round-trips dropped. Not eliminated — but the first-pass questions got sharper because reviewers had more to work with upfront.

What This Week Confirmed

Speed gains from AI are real, but they're not random. They show up consistently in two places: mechanical transformation (restructuring code, reformatting text) and blank-page problems (starting a description, a doc, a test). The judgment layer still belongs to the engineer. That's not a caveat — that's the workflow.

Week 8 down. The pattern is getting clearer every Friday.

I break down one workflow like this every week in The AI Leverage Weekly — practical, no fluff, free. It's the one email I actually write for engineers who want to work smarter without the hype. Subscribe: https://theaileverageweekly.beehiiv.com/subscribe?utm_source=devto&utm_medium=article&utm_campaign=roundup_w8