How 3 Java Champions are using GitHub Copilot for increased productivity

GitHub published a blog post in its AI & ML series gathering tips from three Java Champions, each with more than a decade of Java experience, on using GitHub Copilot in production and open source work. The framing is that fast code generation does not by itself solve the hardest engineering problems: developers still supply context, validate output, protect maintainability and stay accountable for what ships. What follows are the named contributors, their concrete examples and the practices they describe.
Vlad Mihalcea, founder of Hypersistence and author of High-Performance Java Persistence, works on Hypersistence Utils and Hypersistence Optimizer. He described a long-postponed web application that shows Optimizer event history, so teams can see how performance work is actually affecting application performance. What he estimated at weeks of effort took two days with Copilot, and the app is scheduled to ship in the next version of Hypersistence Optimizer. He runs Copilot inside IntelliJ IDEA, where it is convenient during source review and debug sessions, and says it speeds up investigations and prototyping. He still reviews both the plans and the generated code, and notes that the output is sometimes unsatisfactory and needs his intervention.
His techniques center on making generated text and code resemble his own work. One skill is trained on his voice from 65 published articles, so documentation drafts read like his blog rather than a language model, drawing on 12 years of blogging about principles such as expressiveness, testability and simplicity. He developed further skills that steer AI toward existing codebase conventions, citing the principle of least astonishment, on the view that maintenance consumes more time than initial creation. He also varies the model by task: Opus with the 1 million token context window at Extra High or Max reasoning for deep thinking, and lighter, lower-token models for reports, release notes or documentation checks.
Patrick Baumgartner, a software crafter at 42talents with over 20 years of Java experience, coaches teams without pausing normal business. He described three projects. The first was migrating about 900 teaching slides across 15 presentations from Reveal.js to Marp, many containing code, animations and live-editable JavaScript examples that had already been refined with students. The work began with analysis, a target definition, a migration plan, constraints and examples of what had to stay intact, then moved from structural changes to details. Verification was the core: screenshots of both versions compared visually and algorithmically, a small app running both decks side by side for manual animation checks, and comparisons of generated PDFs. The migration took roughly eight to ten hours against an estimated week or more by hand. His stated principle: delegate execution, not responsibility.
The second project was a terminal UI application built with Tamboui that worked under WSL on Linux but, when a real user ran it in Windows CMD or PowerShell, ignored the Enter key in some forms. Copilot helped turn the feedback into reproducible UI tests: it drove the application through tmux on Linux, which led him to psmux for equivalent PowerShell flows. They reproduced the bug, patched locally, traced the root cause into Tamboui and submitted an upstream fix that was merged. When test descriptions outgrew the available context, Copilot proposed a compact arrow notation (step to step to input) carrying extra detail only where critical, a representation optimized for the agent rather than for humans while keeping semantics.
The third was a children's math learning app covering addition, subtraction, multiplication, division and division with remainder for grades one through four. Research with Copilot into spaced repetition, immediate feedback, adaptive difficulty and gamification without dark patterns challenged his plan for a universal app; he scoped it to the Swiss curriculum instead, shipping in German, French, Italian and English with cooperative and competitive two-player modes and spaced repetition after mistakes. It has no backend, keeping children's data on the device and allowing hosting on GitHub Pages.
Baumgartner says he has largely abandoned classic autocomplete for agentic workflows, using VS Code for tight visual feedback loops and the terminal coding agent for larger execution tasks. On a roughly four-hour trip to JavaLand he steered an agent from his phone and arrived with a working Maven plugin that optimizes which Spring Boot auto-configurations load, plus a CI pipeline, performance tests and reports. He uses MCP servers such as Context7 for documentation and an integration with his Logseq knowledge base, but prefers a good CLI when one exists. His advice: stop one-shot prompting, ask the agent to interview you to surface implicit assumptions, demand a plan that shows assumptions, verification gates and architectural fit, define task-specific gates so done means evidence rather than a model's claim, turn lessons into reusable Agent Skills (including a maintainer-audit skill checking docs, CI, Dependabot, CodeQL and cleanup), and wrap the probabilistic model in deterministic tools such as Checkstyle, PMD and Spring Modulith checks run from hooks in a change, check, fix loop. He also rejected an early Spring Boot 3.5 and 4.x dual approach once the older path proved unnecessary, calling detours expensive.
Frank Delporte, senior technical writer at Azul and maintainer of Foojay.io with over 15 years of Java experience, describes Copilot as a companion for evaluating pull requests and investigating bug reports. He cites the Pi4J library, which lets Java applications drive electronic components on the Raspberry Pi; version 4, released in February 2026, adds a plugin built on the Foreign Function & Memory API.
Across the accounts, a pattern emerges. The time savings are real but concentrated in tedious, well-specified work: migrations, test scaffolding, documentation, boilerplate. The gains come with an upfront cost in context and planning that the contributors treat as non-negotiable, and a verification cost that often rivals the generation itself. This suggests the leverage sits in the harness around the model, reusable skills, deterministic checks and explicit gates, rather than the model alone, and that teams investing in that harness will see more durable returns than teams chasing prompt tricks.
Why it matters: Java teams evaluating Copilot get a practical playbook from experienced practitioners rather than benchmarks: context first, plans reviewed, verification gates, reusable skills and deterministic quality tooling around the agent. The examples suggest productivity gains show up fastest in migration and maintenance work, which is where many teams carry the most backlog. My inference is that the differentiator will be process and tooling discipline, not model choice.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Three Java Champions share how they use GitHub Copilot to solve problems, accelerate development, and keep humans responsible for the results. The post How 3 Java Champions are using GitHub Copilot for increased productivity appeared first on The GitHub Blog .