As AI coding assistants become more capable, the relationship between developers and their tools is evolving beyond a simple autocomplete. AI won’t replace developers, but we need to rethink the way developers work with AI to maximize speed and quality.
One highly effective approach is what I call the “head chef” model. Much like a head chef doesn’t chop every vegetable or stir every pot, developers will no longer be writing most of their own code. Instead, they’ll manage a team of AI “sous chefs” that take care of the implementation while a human manages the overall design and quality control.
From coder to system architect
This model changes the work developers do. Instead of spending hours writing code and debugging syntax errors, developers decompose problems into clear tasks, evaluate architectural trade-offs, and verify that AI-generated outputs meet production needs. Developers are the decision-makers, responsible for vision, judgment, and verification, while AI is the lightning-fast assistant doing most of the actual coding.
This division of labor creates what’s known as the FAAFO approach — fast, ambitious, autonomous, fun, optionality. It frees up developers to explore multiple implementation paths, prototype different solutions in parallel, then apply their judgment to validate and merge the most promising elements.
Context management is everything
For this model to work, context engineering plays a critical role. The quality of an AI system’s output correlates directly to the quality of the input. That means learning to curate the right context, like code snippets, documentation, error messages, or architectural constraints, and feeding that context into the AI system in digestible chunks.
If you get results that are inaccurate, it’s often because you provided either too little context, leading to hallucinations or generic suggestions, or overwhelmed the AI with irrelevant data. The key is modular thinking by breaking down your codebase and tasks into clear, manageable parts that the AI can process.
I think of this as a “clipboard” problem. Whatever goes on your clipboard before you paste it into an AI prompt determines how good the outputs will be. The best developers develop an instinct for what context they need to include and what to leave out.
A real-world example: Building Kafka pipelines
For data streaming applications, this model becomes particularly powerful. Imagine you’re building a complex, real-time pipeline with Apache Kafka and/or Apache Flink. An AI assistant can generate Flink jobs that process the streaming data, suggest the best configurations for throughput and latency, write comprehensive unit tests for stateful operators, and even propose schema changes to match how your data model changes.