REPL Driven Development as an analogy for proper AI Driven Development

and a guide from a by-gone era

I recently "discovered" using an LLM to code with, and was looking forward to doing some "REPL Driven Development" on it. Unfortunately, I discovered that all of the vibe-coded vim plugins just use vim as a (terrible) multiplexer for the Claude Code CLI program. The Emacs plugin claude-code-ide.el does it right by having text selection and other send-to-repl-like actions like a proper Lisp IDE does.

This was merely a lisp thing to do because the parentheses clearly define selection to send. That's no longer an issue, the LLM can filter out the garbage in a way that a GOFAI symbolic rule-based pattern matcher can't.

The command line program likes to use diffs to edit files, and that is reasonable, but vastly inferior to REPL Driven Development from a scratch file. Below are some of my notes from 2019 for more context on using the lisp REPL. With RDD, you will never type something directly into a REPL again. "Never type into the REPL".

It's nice using grep to lookup some code I might have written 5 years ago that would have been ephemeral and lost at the REPL.

REPL driven development

Stuard Halloway's 2017 talk at Chicago Clojure:

vimeo
slide notes

DON'T type into the repl. Type into a file that you can search later.
Use repl-interaction commands (like insert-in-repl, minibuffer, etc.).
Have a plan to avoid wasting time (since repls are fun).
Know all the jump-to-documentation kinds of commands. Searching for a regex,
jumping to the source of a thing, etc.
Analysing data carefully at the repl (don't just stare at stacktraces).
Debugging at O(log n)
Custom REPLs (especially debuggers/inspectors with breakpoints, etc.)
Do use the debugger.

MOST important keystrokes

Execute a thing in the minibuffer
Tracing functions (or a short function to execute in the minibuffer like =trace=)
Sending to repl from behind the point
Sending to repl from in front of the point
Sending the whole top-level expression at point to the repl
Jump to the minibuffer (not for typing! just to see big output)
A keystroke to jump back to the source file (or something like windmove)
all the documentation commands (search, jump to source, doc-at-point, etc.)
Execute something and replace it with the output/write the output as

a comment in the scratch buffer.

in: articles