← Back to feed
Constellation Building

Watching a Coding Loop Grow Up

A coding loop is a tool that produces code. Once you can ship the tool, the next question is whether the tool can ship improvements to itself. This arc tracks that journey across phases: build → harden → recurse → escape.

2 signals
—

A coding loop isn't supposed to be the product. It's supposed to be the way you ship products. Mine became a product anyway, because the moment a tool can produce its own work, you start asking whether it can produce improvements to itself. This arc is what happened when I asked that question and tried to answer it in public.

the cost problem
BUILD
Hardening the Ralph Loop: The 68K-Token Tax in My Overnight Coding Agent

My first overnight Claude Code loop produced real commits and still failed: it burned quota faster than doing the work by hand. The real leak was the 68K-token runtime tax hiding before my prompt.

The cost post ended with a destination — bypass the interactive runtime, talk to the SDK directly, give the bash a name (Gantry) and a successor. That destination is still where I'm going. But there was a stretch in between where the bash loop existed, worked well enough to harden, and could be pointed at the loop itself. The recursion happened first because the architecture had to.

the recursion
BUILD
Watching the Loop Forge Itself: A Demo Repo That Refactored Its Own Loop

I pointed my Ralph Loop at its own script to ship a public demo. The useful part was not the recursion — it was every decision I made while the loop made the git log lie, burned $4 on an impossible spec, and taught me which corner I should have cut instead.

The two-track repo taught me what the bugs looked like in the open. The chapter convention scaled. The 3-layer prompt composition let BMAD updates flow in without a fork. But the runtime tax from the first signal was still there — every invocation still loaded the skills, the MCP servers, the plugins. The bash wrapper couldn't fix that. The next signal is about what does.