AI Coding Tools

These essays argue that AI coding tools are not just better autocomplete. Claude Code and The Rise of CLI opens with the numbers that made the question impossible to dismiss: Claude Code shipped on February 24, 2025, and by November had crossed $1 billion in annualized revenue — “ChatGPT needed eleven months to reach a comparable milestone. Slack needed four years.” The puzzle the essay sets out to solve is why, with decades of refined IDEs available, developers chose to “abandon their polished graphical editors to type instructions into a terminal and watch an AI respond in plain text.”

The answer the essay reaches for is not about AI but about Unix. The CLI’s design philosophy — “small programs that do one thing well, connected by pipes,” with text as the universal contract — turns out to be exactly the substrate an LLM agent can compose against. The same properties that made Unix tools intimidating to newcomers ("grep… a name that means nothing until you learn it stands for Global Regular Expression Print"; awk “named not after what it did but after its creators”) are precisely what makes them legible to a model that has read billions of tokens of shell. The graphical revolution rescued users from the terminal by making files things you could see and drag — but agents do not need to see. The CLI never went away; it was waiting for a different kind of operator.

Claude Skills, Commands, Agents toward a Unified Mission adds the next layer. The opening analogy — explaining your dietary restrictions at every restaurant, or carrying separate cards for Japanese, Italian, and bakeries — frames the problem of repeated-prompt overhead. Skills, slash commands, and agents are “all essentially just prompts under the hood — markdown files containing instructions,” but they differ in who initiates them and how they load. Skills make Anthropic’s bet on intent matching explicit: “If you say ‘polish:’ followed by some text, Claude knows you’re not talking about nail polish.” Once that property holds, instructions stop being one-off prompts and start behaving more like reusable software components. The architectural principle the essay foregrounds is progressive disclosure: Claude loads the skill index by default, but reaches for the full SKILL.md only when a request triggers it.

How We Build Software in the Age of AI extends the argument outside the IDE. Software has historically been built for humans, with humans serving as the integration layer threading disconnected tools by hand: “the human as integration layer — threading disconnected tools by hand.” The IDE, the essay observes, “doesn’t spin up the development server when you finish writing. It doesn’t open a browser to check how your changes look… You switch windows. You type the commands. You are the pipe.” Once an agent can play that role, the unstated assumptions — that software is built for people, that humans are the connective tissue — start to wobble. Coding tools are where the new interface model becomes easiest to see in public: software built for agents, with CLIs, skills, and harnesses, rather than software that assumes a human clicking through a GUI.

Taken together, the three essays make AI coding tools the leading edge of a more general shift. The CLI is back, but as the substrate for a new interface era; skills turn instructions into portable components; and the software ecosystem itself begins to reorganize around what an agent can actually do.