Large language models have opened up completely new possibilities in recent years under the banner of artificial intelligence (AI). Accordingly, all stakeholders are making great efforts to adapt to this technological development. Software development is particularly affected, as code is a textual medium and generative AI is particularly good at generating text and thus code. However, the term „AI coding“ encompasses more than just the application of AI-based technologies—at least if one takes the topic seriously.
What changes will artificial intelligence bring to software development?
„Not reinventing the wheel“ is a fundamental principle that is often heard, especially in software development. Nevertheless, developers deal with lines of code every day that exist in countless other projects in exactly the same or extremely similar form. Large language models are the key technology for training artificial intelligence with precisely these repetitive patterns and then reproducing them as needed. Computers thus gain the ability to generate „boilerplate“ code and even adapt it to the current context. In the past, specific instructions had to be defined for each automated adjustment in the code, which could then only be applied in a very strict context. A specific algorithm produced a predictable result for a specific input. Deviations led to exception errors. With large language models, we now have a functional building block that delivers a result for any input. The new art is to formulate the appropriate input to achieve the desired result. In the context of artificial intelligence, this input is understood as a prompt.
„AI coding defines new disciplines for software developers: prompt engineering“
During the rapid development of large language models over the past few years, the term „prompt engineering“ has been both defined as a completely new profession and ridiculed. In the end, the whole thing has settled down to the point where anyone who wants to work effectively with artificial intelligence, regardless of their background, has to deal with creating effective prompts. Anyone who wants to develop software with AI will see that the activity is shifting from writing code to writing and developing prompts. Even though tools such as AI coding IDEs automate a large part of prompt generation, understanding how the prompt works is a key factor in the quality of the output. Only by understanding how it works can you effectively adapt AI to your own requirements and processes.
The term „prompt engineering“ and thus also „AI coding“ encompasses the following additional skills:
- Prompt Management – Prompts are the driving force when working with LLMs. The quality of the prompt determines the result. That's why prompts must be treated and managed as a resource. Individual projects require individual prompts that need to be created, shared with colleagues, and integrated into tools. Using protocols such as MCP Prompts Prompts can be integrated into tools.
- AI Scripting – Since problems cannot always be solved with the same prompt, there must be ways to influence the prompt flexibly and sustainably (not manually). Many AI tools use a template engine such as. Jinja, to define flexible prompts. Assembling prompts using scripting languages is also a powerful method. There are many frameworks and libraries that specialize in the creation of AI-supported programs, including specialized programming languages (e.g.,. BAML).
- Context Engineering – AI-driven processes do not take place in isolation. The task of the „AI coder“ is to ensure that the prompt is built with up-to-date and relevant information. Technologies such as RAG, MCP, Scripting, template engines, etc. are ways to design prompts depending on the environment.
- Tool Integration Existing tools and processes are not rendered obsolete by AI, but rather complement each other. One task of „AI coding“ is to implement appropriate integrations. This can be done via APIs, plugins (e.g., for IDEs), or processes. Protocols such as MCP, ACP, A2A offer standardized recipes for integration. Workflow engines such as Dify, n8n or sim come with a large number of connectors.
„AI Coding – Toolbox
To put theory into practice, software developers have logically created software solutions that integrate and automate our familiar tools such as IDEs, editors, pipelines, etc. with new technologies such as LLMs, prompts, etc. Here, too, there are various solutions and approaches that you should understand as an „AI coder“ and integrate into your own process. The solutions can be classified, for example, according to the size of the feedback loop (from fast and small to large and slow) and the proximity to the code (which is usually represented by an editor). The spectrum ranges from suggestions in the editor for the next line of code to coding agents that run somewhere in the cloud and generate code according to the principle of Vibe Codings Hide completely from the user.
- In the editor
- Inline code completion
- Next Edit code completion
- Inline edit
- Inline code generation
- In the IDE
- Fix with AI
- Ask about code
- Chat with code
- Code review
- In the same environment
- Agent in terminal
- Agent Chat
- On a server
- Agent on server
- Workflow engine
- Agent in pipeline
- Regardless of the code
- Vibe coding service
The further down the list we are, the further away we are from the code. Vibe coding, completely without seeing the code, is an extreme at the end of the scale that makes little sense for experienced software developers, as their greatest strength is the use of code as a powerful tool. Accordingly, code can also be a potent tool for artificial intelligence, such as. The latest developments regarding MCP and „tool calls“ shows.
Leave nothing to chance
The exciting question is how much we want to leave to our artificial colleagues. Working with AI is often like operating a slot machine with completely unexpected results and surprises. This is where professional „AI coders“ differ from AI users. Just because a task can be performed by AI does not mean that it must automatically be performed by AI.
Here is an example:
Use the provided API to get the issue informations Implement the issue
In this variant, we leave it to the AI to retrieve the information. This requires additional contextual information, is more difficult to understand, and has the potential to fail.
{{ curl http://issue-tracker.com/issues/1234 }}
Implement the issue
In this variant, we retain full control over the retrieval and the information it contains. We receive a prompt that is clearly focused on the main problem, save ourselves the trouble of providing the LLM with additional tools, and can effectively reproduce the result.
Conclusion: AI as an amplifier, not a replacement
AI makes software developers not superfluous – it unfolds its potential only in cooperation with a AI Coder.
Anyone who takes AI coding seriously,
- leads AI (instead of being guided),
- retains control (instead of leaving results to chance),
- uses AI as a tool (rather than as a black box).
→ The future of software development belongs to those who both master: code and AI.
