Installation Docs 'For Humans': Just Paste This Into Your LLM Agent

Level 1: The Recipe That Says "Don't Cook"

Imagine opening a cookbook to a recipe titled "For Home Cooks," and the only instruction is: "Hand this book to your robot chef. Honestly, you'd just burn it." Then you notice the next chapter is titled "For Robot Chefs." The joke is that the instructions for people are really instructions to step aside and let the machine do it — and the book isn't even apologetic about it. It's funny because it's a little insulting, a little true, and a sign of how fast the kitchen has changed.

Level 2: Agents, Prompts, and Raw Files

Some terms doing the heavy lifting here:

• LLM agent — an AI assistant (like Claude Code or Cursor) that doesn't just chat: it can run shell commands, edit files, and install software on your machine on your behalf.
• Prompt — the text instruction you give the AI. In this README, the "installation command" is literally an English sentence, not a shell command.
• Raw GitHub URL — a link serving a file's plain contents instead of the GitHub web page around it. Agents fetch it to read the instructions directly. Note it points at the dev branch, so the instructions can change at any moment.
• Fat-fingering — typing errors: the classic junior-dev rite of passage where one wrong character in a config file costs an afternoon.

The early-career takeaway: when you delegate setup to an agent, you trade your typos for its judgment. The install might go perfectly, or the agent might cheerfully resolve an error by doing something you'd never approve. "Read the Installation Guide" remains good advice even if the README itself is sarcastic about it — you can't review what the agent did if you never knew what it was supposed to do.

Level 3: The README Has Opinions About You

This is a real installation section from a dark-mode GitHub README, and the heading hierarchy is the whole joke. Under "For Humans", the instructions do not tell a human how to install anything. They say:

Copy and paste this prompt to your LLM agent (Claude Code, AmpCode, Cursor, etc.):

followed by a code block containing a one-sentence prompt — "Install and configure [project] by following the instructions here:" — pointing at a raw installation.md file on the repo's dev branch. The actual human-readable path is demoted to an afterthought:

Or read the Installation Guide, but seriously, let an agent do it. Humans fat-finger configs.

And cut off at the bottom of the screenshot, the next heading: "For LLM Agents" — meaning the section labeled for humans is agent instructions, and there's a separate section for agents. The taxonomy has inverted. Documentation used to be written for people, with machine-readable bits bolted on; here the human is a transport layer whose only job is Ctrl+C, Ctrl+V.

What makes this sting for anyone who's lived through prior installation eras is the lineage. We went from hand-edited Makefiles, to ./configure && make install, to package managers, to curl | sudo sh one-liners, and now to natural-language prompts as the install command. Each step traded inspectability for convenience. This one trades it for a system that downloads instructions from a moving target (refs/heads/dev — not even a tagged release) and then improvises with whatever permissions your agent has. Prompt-injection researchers call docs like this an attack surface: whoever can edit that markdown file can now steer an agent that you invited into your shell. It's the supply-chain problem from curl | sh, except the payload is interpreted by a model that might also helpfully "fix" your dotfiles along the way.

And yet — the maintainer isn't entirely wrong, which is why this is satire-adjacent rather than satire. Humans genuinely do fat-finger configs; transposed ports, stray tabs in YAML, and missed prerequisite steps account for an enormous share of "your software is broken" issues. Agent-first docs are a rational response to a support burden. The unsettling part is the casual openness: the industry quietly decided humans are the unreliable component, and this README just said it out loud in the DeveloperExperience equivalent of a hot mic moment. The channel's caption — "Anyway, we've got a new meta" — frames it exactly right: this isn't one weird repo, it's the new equilibrium.