Coding with Copilot – New models, new features in check

Software engineering is undergoing fundamental change. Not gradually, not theoretically – but noticeably in everyday working life. The new language models such as Claude Sonnet 4, GPT-4o or Gemini 2.5 Pro are changing the way code is created. They automate the repetitive and create space for the essentials. Combined with GitHub Copilot, they create workflows that do more than just save time: They redefine quality standards.

But hand on heart: how much of it is hype and what really works? Our developers put the new models through their paces. The results? Surprisingly practical.

Particularly fast and efficient, it is suitable for everyday tasks such as commit analyses or the generation of activity reports.

Strong in the multimodal area – ideal for processing images, UI mockups or visual documentation.

The context king among the current models. With a window of up to 2 million tokens, even huge code bases or entire projects can be analyzed at once.

Not yet integrated into GitHub Copilot, but promising in terms of long-term planning and logical coherence.

All in all, these models offer a strong arsenal for use in productive development environments – provided they are used in a targeted and context-aware manner.

A question that has kept our developers busy: Can booking entries be generated automatically from commit data? The idea behind it: If developers commit regularly anyway, it should be possible to use this information to generate activity records.

This is exactly what was tested – with success. Claude Sonnet 4 was instructed by prompt to extract all commits from Git repositories within a certain date range. The answer came in the form of structured JSON files with commit descriptions, changed files and timestamps.

The quality of the results was impressive: the automatically generated texts matched the actual work process with astonishing precision. In many cases, the result was even better structured than manually generated reports.

Compared to earlier models, Claude showed clear advantages Here

• Better error correction:
  Even if an initial version did not work, repeated requests often led to a usable result.
• More initiative:
  Claude added functions that were not even thought of – such as additional log data or help functions.
• Contextual understanding:
  Existing API documentation did not have to be handed over separately – Claude already “knew” how Azure DevOps works.

Alongside the new models, GitHub Copilot has also undergone significant development. Particularly relevant:

• Instructions:
  File and project-related control commands such as “always generate documentation”, “always reply in German” or “no praise in the reply” can be stored centrally.
• Prompts:
  Reusable task descriptions for typical tasks – e.g. conversion from CommonJS to ESM, code reviews.
• File-specific rules:
  e.g: Always prefer grid layout for CSS files, do not use px values – simply use *.instructions.md.

Copilot is also a valuable assistant for code reviews. The review function recognizes anti-patterns, such as hardcoding an API key in the source code – and automatically suggests improvements such as .env variables.

The reviews are file-based or via a tool button for uncommitted changes.

A particular strength is demonstrated with larger code bases: Thanks to so-called agentic workflows, Copilot can process several files in parallel. One example:

The prompt read: “Search all .js files for .then() constructs and replace them with async/await.”

Copilot then generated a bash logic, extracted the affected files and performed the conversion – all with minimal interaction. A prime example of productive automation.

The combination of modern language models and GitHub Copilot is changing software development for good. From intelligent time recording and automated reviews to complete refactorings – the possibilities are real and practical.

The key to success? Curiosity and targeted use. Those who understand the tools and use their strengths gain time for what really counts: well thought-out architecture, creative solutions and better software.