Using LLM in the shebang line of a script

Why This Matters You might dismiss this as a neat technical trick, but it touches on a deeper evolution in programming paradigms. Traditionally, the shebang line (#!/...) tells the operating system: "This file is a script; please run it with /usr/bin/python or /bin/bash." Simon Willison's demonstration shatters this convention: now you can write a file where the first line is #!/usr/bin/env -S llm -f, and from the second line onward, describe what you want in plain English, like "Generate an SVG of a pelican riding a bicycle." Save it, make it executable, run it, and the system invokes an LLM tool to generate that image. This is no longer "write code for a machine to execute," but "write a description for an AI to understand and execute." What It Changes

1. A Revolution in Interaction: The most immediate change is lowering the barrier to creating a "program" to its absolute minimum. You don't need to learn Python syntax or figure out how to call an API; you just need to describe your need clearly in language. It's like issuing a "natural language command line" to your computer. 2. A Leap from Procedural to Declarative: Traditional programming is "procedural"—you must tell the computer step-by-step how to do something (define a variable, loop, then check a condition). This approach is "declarative"—you only tell the computer what you want ("generate an SVG of a pelican riding a bicycle"). The specifics of how (which model to call, how to parse, how to generate) are handled behind the scenes by the LLM toolchain. 3. Seamless Integration of Tool Use: The article shows more powerful patterns: using the -T option to call tools (like llm_time to get the current time) or defining Python functions via YAML templates (like add, multiply). This means your "natural language script" can break free from the confines of pure text generation and interact with the real world (fetch time, perform calculations, query databases). This effectively packages the capabilities of an AI Agent within the most traditional form of a script. How This Relates to You For developers, this offers a new mindset for rapid prototyping and task automation. You can use it to quickly create "one-off" tools for processing text, generating content, or answering questions without setting up a full application framework. For example, create a script that automatically summarizes the latest articles on your blog daily, or queries a database and generates a report based on a natural language instruction. The more profound impact lies in its foreshadowing of the generalization of "programming." In the future, collaborating with computers may not be limited to writing code with strict syntax. The ability to describe problems and requirements clearly and accurately could become a core "programming" skill. This simultaneously raises the bar for LLMs in terms of instruction-following ability and tool-use reliability—how robust your "natural language program" is depends entirely on the underlying AI's "understanding" and "execution" capabilities. A Counter-Intuitive Perspective Many believe AI will "replace" programmers, but this example suggests another possibility: AI is democratizing the ability to "program" for more people while also redefining what a programmer does. Future "advanced programming" may not be about mastering complex algorithms, but about mastering how to design precise prompts, orchestrate reliable tool flows, and validate AI outputs. It's a shift from being a "person who writes code" to a "person who designs intent and validates results." In summary, this seemingly simple trick acts like a prism, refracting the future landscape of human-computer interaction in the AI era: more natural, more direct, and more focused on intent than implementation.