An office with a dustbin and windows; a desk with documents, files, and folders. Pushbuttons to trigger actions and menus to offer choice. An inbox for mail; messages and announcements posted on a wall. Portals as gateways, platforms as infrastructure, and channels as paths. Streams of thought, scrolls of memory. The metaphors of digital interface design draw from the analogue existence of its end users and the experiences with which they are familiar. The most relatable analogy of all, a chat, thus becomes the ultimate cue to a lived knowledge that has made it through multiple porous informational thresholds. As with any metaphor, the chat, too, facilitates a disguise—here, the power and complexity of AI tools are concealed behind the most powerful and complex of human means: language.
The instant gratification of contemporary chatbots—the idle small talk accomplishing major technical endeavours—distracts from the computational technologies these conversations exploit. Human users communicate with digital machines through conversational questions, reasonably expecting reasonable answers. Generative AI may look like a dialogical search engine or a verbose digital assistant, but information is created, not found in these systems.
The stormy pace of that discontinuous thing called ‘progress’ expects interfaces after AI to be the final form of UX design. Dialogue systems such as the chat window of a large language model are confined to turn-based conversations with little or no capacity to handle multiple modes of human communication and experience and are thus perceived as limited. Arguably, this is an interface with just one overriding profile—an interface with one silhouette and shadow.
A cursor is blinking. It signals, ‘I am ready. Tell me what to do. Tell me where to go.’ Command-line interfaces (CLIs) are remainders from the neon-green past of computing. Although fundamentally distinct in origin, evolution, and scope, the chat box of contemporary AI-driven chatbots evokes this bygone dimension of computation, reprising some of its linear and textual features. CLIs were introduced in the 1960s, replacing punch cards to control automated machines. By the 1980s, when friendlier, more intuitive graphical user interfaces (GUIs) took over, they had fallen out of fashion and favour. The operational principle of a command line is simple: users type an instruction directly to the machine. No icons, no windows, no pointing devices; just text in and text out. When the computer responds, success messages are minimal. You know it works because it has worked. Press Enter, and the command is executed. No guessing is allowed; the machine follows the command exactly as given, then pauses and waits for the next instruction. A shell—the interpreter processing the command—shields the operating system’s core, the engine that makes things happen.
Prompting is common practice today, when even university degrees are marketing themselves as training grounds to upskill the prompt engineers the automated workplace will need. Everybody knows that, for the new AI, a prompt is a natural language input given to the generative model for it to produce certain outputs. A prompt is a clear direction; it can include contextual information and background examples or set parameters, formats, and limitations. For a CLI, on the other hand, a prompt is something else: not a request from the human user, but a textual display (for example, a symbol like ‘$’, ‘#', or ‘>’) signposting that the machine is ready to accept orders from the user.
While a prompt on the command line is a marker for an input, visually confirming the machine’s status, displaying fixed information such as a username, working directory and system state, for a generative model, it is a human-originated instruction through which to elicit a desired machine response. Two different modes of compelling and two different modes of interacting and interfacing, each with its own set of distinct directionalities, purposes, and functions, yet both functioning as effective reminders to do something, engaging and elevating machine agency. A prompt given to an AI model does a lot, shaping and directing the model’s output, influencing the programme’s behaviour, and modifying itself through interaction. The prompt on a CLI, conversely, just waits. It shows information without affecting how commands work. A prompt in generative AI is (somewhat) interpretative, while a CLI prompt is mostly literal, with the shell—that boundary between an inside and an outside—executing precise programming instructions and a cursor indicating where to type them.
The seeming passivity of these CLI exchanges is, however, deceptive. Yes, on a command line, waiting is perpetual. Yes, it is also meaningless. Nevertheless, such waiting is purposeful. While, technically speaking, these computational mechanisms will not act per se but only respond to action, they are not there just to attend, serve, and oblige. As with anything in computing, between input and output there is always processing. On a command line, this processing is what is solicited, invoked by the flickering prompt, the request for a command. A superior semiotician, Charles Sanders Peirce, once proposed that signification does not belong to the human but that the human is itself a sign. Then, the prompt in the shell identifies the space where a human input is such a sign, where a symbolic something is presented to stand for a performative something else, namely, in the processual execution of an idea via the ideation of a world made out of that input sign.
AI interfaces may soon achieve the seamless combination of text, images, voice and gestures while building context-aware forms of ambient computing for immersive interaction and augmented reality. Interfaces after AI will likely be shape-shifting adaptations of the user’s every need, maximal AI expressions leveraged through any internet-connected medium or any source of information. With the state of anticipation of a prompt in the shell being an end in itself, however, a CLI demonstrates how interfacing is both connecting and separating, sharing a margin that does not have to expand and inflate but can also compress and deflate.
On a command line, syntax is semantics and brain is brawn. On a command line, to interact with computation is to compute. The prompt of a command line reminds us that any fantasy of unmediated communication, pre- or post-AI, is an illusion. It does not crush a dream, though, but rather allows distinct aspirations for human-machine relations. Waiting in line, like a prompt in the shell, reveals computation as a space in which everything is actualised thought. Despite many differences, any AI chat window encloses these command-line truths with its blinking cursor. It is not merely a passive receiver of requests. It is close enough to but also sufficiently distinct from its human input to accept that as the representation that connects and separates—or, better, interfaces—ideation and execution.