Goo to Structure

There’s a little Python project called nlsh that does one thing:

You type “whats in my disk” and it outputs df -h.

That’s it. Natural language in, shell command out. Human goo → machine structure.

The LLM sits in the middle as a translator. It takes your fuzzy intent—misspelled, vague, context-dependent—and converts it into the precise syntax the machine needs.

"whats in my disk"  →  df -h
"delete that file"  →  rm ./output.log
"go back"           →  cd ..

This is what Claude Code does too. You say vague shit, it emits tool calls with exact parameters. You describe what you want, it writes the grep command you’d never remember.

The irony: LLMs are famously “fuzzy”—probabilistic, hallucination-prone, unreliable with precise tasks. And yet their killer app is converting fuzziness into precision.

They’re not good at being precise. They’re good at understanding imprecision.

The goo-to-structure conversion is the interface layer we’ve been missing. Computers have always demanded exact syntax. Now there’s a translator in the middle that speaks both languages.

Human goo in, machine structure out.