Three Java Champions share GitHub Copilot workflows: weeks-long tasks done in two days

GitHub Blog AI/ML · rss · 2026-10-21

Overall stance

The GitHub Blog interviewed three Java Champions with 15–20+ years of experience (Vlad Mihalcea, Patrick Baumgartner, Frank Delporte) about using Copilot in production and open source. Their shared view: AI generates code fast, but developers still supply context, validate results, protect maintainability, and stay accountable.

Vlad Mihalcea (Hypersistence founder)

A web app he'd postponed for weeks took two days with Copilot, shipping in the next Hypersistence Optimizer release.
Uses it in IntelliJ IDEA; still reviews plans and code himself.
Custom skills: one captures his voice from 65 published articles so drafts sound like his blog; another makes AI mimic existing code practices per the principle of least astonishment.
Switches models by task: deep thinking uses Opus with 1M context and max reasoning; lighter tasks (reports, release notes) use cheaper models.

Patrick Baumgartner (42talents)

Migrated 900 teaching slides across 15 decks from Reveal.js to Marp: analyze, plan, set constraints, then big-to-small. The key was verification — running both decks side by side, comparing screenshots algorithmically, diffing PDFs: "the agent didn't just say done; we had evidence." Took 8–10 hours vs. a week+ manually.
A Windows Enter-key bug in a terminal UI app: turned user feedback into reproducible UI tests (tmux on Linux, psmux in PowerShell), traced it upstream into Tamboui, and got the fix merged.
When test descriptions blew the context window, he and Copilot co-designed a compact arrow notation (step → step → input).
Core principle: delegate execution, not responsibility.

Original post →

Three Java Champions share GitHub Copilot workflows: weeks-long tasks done in two days

More from coding & agent