There’s a black screen. A blinking cursor. Nothing is happening, and nothing will happen … not until you convert thoughts to keystrokes. All of those taps are minor translation projects from cognition to binary.
Thus was my initial experience with coding — and now with skills and agents and commands defining the substrates for agentic workflows, I’m doing more typing than ever. My first computer science projects were headphones on, terminal up. But in this newly-inspiring LLM landscape that sits before us, the terminal has become even more engrained to me now than then. I have more console tabs open than I do browser tabs. I’m seeing more terminals open in coffee shops.
Crystalizing this for me is that grep has now slipped into my active vernacular as a verb only recently, even though it’s been part of my nerdy milieu for as far back as I can remember. When I initiate a new session with an LLM, one of the frames I use in my own mental model is “What does it need to grep in order to build the context it needs?”
The origin of grep is a Unix command from 1974, a sequence of four letters that unlocked the most powerful aspects of compute. You give it a pattern, a regular expression and the global context to search; it returns the matches. Find the signal, discard the noise. To my mind grep was the inverse of grok"Understand empathically" — an arbitrary formation by Robert A. Heinlein in Stranger in a Strange Land (1961). Said to mean etymologically "to drink." Attained popular use in 1960s–70s counterculture.— etymonline.com
— an invented term from the same era. Rather than seeking outward to pinpoint something — grepping — if one “grokked” something it meant to understand it from the inside out, as if consumed by it.
The terminal used to demand exactly this fluency — pattern, precision, intent. To grep well, you had to grok first. You supplied the needle, you chose the tool; the machine furnished the haystack.
Yet now, the binary can talk back, and it seems like cognition.
It’s still just software
It’s not just a black screen anymore — we’ve gone from an environment of deficit to a marketplace of surplus. But surplus has its own failure mode. The harnesses[A]n agent harness is the software infrastructure that wraps around a large language model (LLM) or AI agent … essentially everything except the LLM itself.— parallel.ai built around LLMs don’t just tolerate vague instructions — they intercept them, asking whether to clarify or proceed. That checkpoint matters: without it, the model follows its best guess down a path you never intended, producing something plausible but wrong. Instead of a crash, you get a confident detour.
Take a typical ongoing project, such as “improve parent communication” when we haven’t yet decided what that means, let alone how to measure it. We are naturally imprecise about what we want, and we still need ways to tease details out. Tools like Claude Code and Antigravity provide us the means to do so.
Schools have long had frameworks for exactly this process as well: SMART goals. Specific, Measurable, Achievable, Relevant, Time-bound. They went viral in presentations and in practices because it forced the translation of intention into something formed enough to act upon.
The terminal used to enforce a similar mechanical precision — you either grokked its language or got nothing back. Now we have software that can detect ambiguity and make a prediction based on the bits and bytes of the given context. Instead of error messages, we get artifacts to modulate — blurring the line between author and editor.
What does this mean for schools?
While barriers have been lifted, what remains however is the underlying requirement to frame the problem and identify outcomes.
Extracting maximum value from AI tools doesn’t require a computer science background — but it does require the kind of discipline that SMART goals were always asking for. The staff who will thrive aren’t necessarily the ones who know the tools. They’re the ones who know what they’re trying to do.
The word grepping didn’t leak into this one person’s vernacular through computer science classes. It’s arrived there now because the concept underneath it — find the relevant thing, bring it in, act on it — turned out to be universal. It was always going to be needed. It just waited, like the cursor, for something powerful enough to put it to use.
The cursor seems to be blinking back at us — but we can still remain in control.