The engineers getting the most out of AI tools aren’t the ones who know every framework. They’re the ones who know where they are in a codebase, how to track what changed, and how to write down what they want.

That’s it. That’s the whole insight.

We’ve spent the last year watching people debate whether AI will replace software engineers. Wrong question. The right question is: what do you actually need to know to direct an AI that writes most of the code?

The answer is three primitives. Everything else, the AI can handle.

The CLI

The command line is how you tell a computer what to do — not through menus or buttons, but through text. If you can navigate a filesystem, read output, and understand roughly what’s happening under the hood, you have control.

AI coding tools — Claude Code, Cursor, GitHub Copilot — all operate here. They read files, run tests, execute commands, and report back. The engineer who can read that output and know whether something went wrong has leverage. The one who can’t is just hoping.

You don’t need to memorise a hundred flags. You need to know that a filesystem is a tree, that processes have input and output, and that when something fails there’s a message that tells you why. That’s sufficient. That’s the floor.

Git

Git is how you track intent over time. Every commit is a decision — what changed, why, and when. A developer who understands the structure of a git repository can follow the history of a codebase like a story.

This matters enormously when working with AI. GitHub data shows 46% of all code written by Copilot users is now AI-generated. If you can’t read a diff, you can’t review what it wrote. If you can’t commit cleanly, you lose the thread of what changed and why.

GitClear’s 2025 research found that AI-assisted code shows a 4x increase in code duplication, and 29% of generated Python code contains potential security weaknesses. The people getting burned are the ones who merged without looking.

Git isn’t optional. It’s the audit trail. It’s how you stay in control when the volume of change accelerates.

Markdown

This is the one people underestimate.

Markdown is how you write structured intent in plain text — headers, paragraphs, code blocks. GitHub’s engineering team now calls it a programming language. Their open-source Spec Kit organises an entire software project around three Markdown files: a spec, a plan, and a task list. Claude Code and Copilot read those files and build.

Visual Studio Magazine put it plainly in February 2026: “Markdown has moved from passive documentation to always-on behavioural guidance for an AI assistant.”

If you can write a clear spec — what this should do, why, what success looks like — an AI can build it. If you can’t, you’ll spend your time arguing with the output instead of shipping.

The gap isn’t where people think it is

Andrej Karpathy coined “vibe coding” in early 2025 — just accepting AI output and iterating on vibes. The Stack Overflow 2025 Developer Survey found 84% of developers are using or planning to use AI tools. But only 29% trust the accuracy of that output, down from 40% the year before.

The gap isn’t technical knowledge. It’s judgment. Senior engineers get more out of AI because they can evaluate what it produced. They can read the diff, follow the logic, spot the clone, and catch the security hole before it ships. Juniors who skip those steps are the ones producing the code that GitClear’s researchers flagged.

The primitives — CLI, git, Markdown — are the foundations of that judgment. You don’t need to know every nuance of React’s render cycle. You do need to know when something changed, why it changed, and whether the change makes sense.

What this means for hiring

Early-stage teams should stop filtering for framework expertise and start filtering for fundamentals literacy. The best candidate isn’t the one who knows the most syntax — it’s the one who can navigate an unfamiliar codebase in ten minutes, write a coherent spec in Markdown, and review a pull request without skipping the diff.

That’s your new bar. The frameworks come and go. These three things compound.

If your team is figuring out how to hire and structure engineering in the AI era, we’ve seen enough patterns to help — reach out at jfsi.io.

Sources: GitHub Copilot — 46% of code is AI-generated · GitClear 2025 AI Code Quality Research · GitHub Blog — Spec-Driven Development: Markdown as a Programming Language · GitHub Blog — Spec Kit open-source toolkit · Visual Studio Magazine — In Agentic AI, It’s All About the Markdown · Stack Overflow 2025 Developer Survey · Wikipedia — Vibe Coding