For most developers, code has always been about certainty. You write instructions, the compiler follows them, and you get predictable output. If something breaks, it's a bug—something you did wrong. But a growing chorus of engineers is pushing back on this mental model, arguing that working with AI is fundamentally different: closer to leading a team than issuing commands to a machine. In an essay published last week on his personal site, developer Allen Bargi draws the comparison explicitly. Traditional code behaves like a compiler—it does exactly what you tell it. People, by contrast, sometimes deliver exactly what you asked for and sometimes deliver something better because they understood your intent. AI, Bargi argues, operates somewhere in between: "The same request can produce a different answer. It can make a useful connection, miss an obvious point, or surprise me with an approach I had not considered." This unpredictability frustrates developers who treat AI tools like enhanced compilers. You expect deterministic output from the same input, and when you don't get it, you blame the tool. But Bargi suggests reframing your relationship with these systems entirely: "It becomes more useful when I treat the interaction as a form of collaboration." This isn't about anthropomorphizing AI or pretending it's sentient—it's about adapting your workflow to match how these tools actually behave.

Building Shared Context

The practical implications are significant. Bargi identifies several practices that improve outcomes: providing examples, offering corrections, and building reusable instructions that reduce misunderstandings over time. The goal is creating what he calls "shared working context"—a foundation that makes the AI's outputs increasingly aligned with how you think and what you actually need. This looks a lot more like onboarding a new team member than writing documentation for a function.

The Skill Shift

Here's where it gets interesting for infrastructure folks: we spent decades learning to tell computers exactly what to do. Syntax, types, error handling—all the machinery of precise instruction. Now Bargi suggests we also need to explain why work matters, what success looks like, and where human judgment should override the system's defaults. That's a leadership skill, not a coding skill. "The technology is new," he writes. "The leadership skills are not."

Key Takeaways

  • AI outputs are inherently less predictable than compiled code—treat them as collaboration, not command execution
  • Shared context and reusable prompts reduce friction more than clever prompting techniques
  • The valuable skill isn't learning to write better prompts; it's learning to express intent clearly
  • Developers who've spent years perfecting precise instructions may need to unlearn some habits

The Bottom Line

This essay is a useful reminder that AI tooling has exposed the limits of the "write exact instructions, get exact output" mental model. If you're treating Copilot or ChatGPT like a more powerful compiler and wondering why results are inconsistent, you might be the bottleneck. The developers who thrive in this new era won't necessarily write better code—they'll communicate intent more clearly.