If you visited my website, you probably saw the Aurora effect behind it: glowing bands of color that drift and blend like a CRT-suffused aurora borealis, with a halo chasing your cursor and a faint particle field that scatters as you move.

What you can’t see from the outside is the process. I built the new UX over a few weeks of pair-programming with an AI coding agent. This is a series about that workflow. It’s about how a static site became a small living system, and what it meant to co-build it with an agent that can generate code faster than I can review it.

Architecture: a theme as a forkable engine

This site runs on Hugo, and its theme is a git submodule pointing at my fork of hello-friend-ng. That split predates the aurora by years, and it turned out to be the single most useful structural decision.

Two repos with clean responsibilities:

  • Main repo (austinatchleyxyz): content, layouts, config.
  • Theme repo: all the custom JavaScript, SCSS, and partials that make the site feel like mine.

Because the theme is a submodule, the main repo never edits theme code directly. Every visual change follows a ritual: commit in the theme repo, then bump the submodule pointer in the main repo. The result is a bite-size changelog that’s readable at a glance:

4e3d18f CRT aurora for landing page
3299ac5 improve styling of CRT aurora. add twinkles
ac12360 use new and improved CRT aurora
63bdbf8 Use TypeScript theme JS toolchain
110584d Fix double-loading of animation scripts and canvases
0d94274 Remove .IsHome guard so aurora/particles load on all pages

This turned out to be an ideal shape for agent-assisted work: each change is small, independently reversible, and carries its own rationale. When the agent and I disagreed, the unit of disagreement was a single commit, not a sprawling diff.

The deploy loop: an idea reaches production in about a minute

The key piece that allowed me to iterate quickly with this whole setup isn’t actually the code generation. It’s the release pipeline. A push to main and I get the result on the live site almost immediately.

The theme repo has a GitHub Actions workflow that runs npm run check on every push: typecheck, lint, prettier, the 49 vitest tests, and a full esbuild build. If any of that fails, the commit is quietly rejected and nothing ships.

The main repo has its own workflow that runs on main: it installs Hugo, does a hugo --minify build, then hugo deploy pushes to S3 and invalidates the CloudFront CDN cache. End to end, from git push to the live site, is on the order of a minute.

This means the agent can try ideas locally with the dev server, and if a change is straightforward enough, the feedback loop is: commit, push, one minute, look at the real site. Cutting the lead time on a new feature down to that length changes how you choose what to try. Experiments that would feel too expensive at a day-long deploy cycle are free at a minute-long one.

The two workflows also enforce the split at the seam between repos: the theme is guarded by its own quality gates, and the main repo has a Hugo-only build that needs no Node toolchain at all because the compiled JavaScript is committed.

The open-source setup

Worth naming the toolchain, because the pricing model shaped quite a few of these decisions. The agent runs through opencode, an open-source terminal coding agent, and OpenRouter, which routes requests to whatever model I want and lets me hot-swap mid-session. No proprietary closed-source harness, no subscription, and if a model goes sideways mid-task I switch to another one in a keystroke.

The model that did most of this work is DeepSeek V4 Flash 0731, which OpenRouter lists at roughly $0.08 per million input tokens and $0.16 per million output, with a 1M-token context. Provider discounts and prompt caching often make the real bill smaller than the list price. For context, a single proprietary subscription runs $20 a month whether or not you touch it. Twelve weeks of this project cost pocket change measured in cents, not a recurring bill.

To put that per-token number in perspective, OpenRouter currently lists Anthropic’s flagship Claude Opus 5 at $5 per million input tokens and $25 per million output. Claude Opus 4.8 lists at the same rate. So DeepSeek comes in around 60x cheaper on input and 150x cheaper on output. Prompt caching narrows the gap a bit, but nowhere near enough to change the calculus.

That gap matters more for agentic work than for anything else, because agents are token-hungry by nature. Every iteration swallows the conversation history, tool calls, repo reads, and file rewrites, often with a large context window that gets re-read across turns. At Opus prices that kind of loop is an operating cost you think about. At DeepSeek prices it’s noise, which changes how you work: I tried experiments constantly and let real failures happen, because the per-attempt price made every try feel free.

Next up: Part 2: Measure Before You Move, about the instrumentation that tells the agent where to spend its effort.