AI-assisted development

Coding assistants now sit in the editor, the terminal and the pull request. They finish the line you are typing, answer questions about a codebase in a chat panel, and in agent mode plan and carry out changes across several files, running commands and tests along the way. All of them rest on large language models, which produce the most likely continuation of the context they are given. That makes them fast and fluent, and it also means fluency is no evidence of correctness.

A working loop

1. Frame the task. State the goal, the constraints, the files involved and an example of the expected result. This is where prompt engineering earns its name.
2. Keep changes small enough to review without losing the thread. An agent that rewrites twenty files hands you a review you will not do properly.
3. Read the diff line by line and run it. If you cannot explain a line, do not merge it.
4. Test independently. Let the assistant draft tests, but choose the edge cases yourself, since a model tends to test the behaviour it just wrote.
5. Check what came along: new dependencies, licence questions, secrets pasted into prompts or files.

GitHub's Copilot exam, GH-300, shows how the vendor itself frames competence here. Its study guide groups the skills into responsible use, the assistant's features in the IDE and command line (including agent mode and the Model Context Protocol), how the tool handles data, prompt and context crafting, productivity uses such as generating tests and refactoring, and privacy controls such as excluding content from the assistant's view. Validating AI output is named as a skill in its own right. For the platform underneath, GitHub Foundations is the entry credential, and Microsoft's free Applied Skills lab on Foundry agents turns the question around: building an agent rather than directing one.

What to keep in human hands

Some code should be written slowly by someone who understands it: authentication and permission checks, cryptography, payment handling, migrations that touch production data, and anything a regulator may ask about. Agentic tools also need boundaries on what they may run, especially in repositories holding credentials. Organisations set these limits through content exclusions and policies; individual developers set them by deciding which tasks to delegate at all.

The beginner's risk is learning to prompt before learning to program. Someone who has never debugged their own loop cannot tell when a generated one is wrong, so use the assistant as a tutor that explains rather than a ghostwriter. GitHub Copilot is the assistant most developers meet first. The skills that keep this work safe live in software testing and software engineering practice; the models themselves are covered under generative AI and LLMs, and the security side under application security.