JERÓNIMO COLÓN III.

Giving AI assistants a safe set of hands on Apple's brand-new container runtime. It's an MCP server — the protocol that lets an AI call real tools — exposing nine tools across three safety levels (read-only by default; opt up to safe, then full), from a system health-check to the full container lifecycle: list, run, inspect, logs, exec, stop, kill, delete. Where the community servers just script the command line, this one talks straight to Apple's background engine through its own Swift client library — faster and sturdier. The long game: a full “Playwright for Apple containers.”

Shipped, live, and serving real traffic — a two-sided mirror for your web browser. The server half reflects back exactly what your browser broadcasts to every site you visit: raw user agent, request headers, IP, and rough location, echoed straight from my backend. The client half surfaces what your browser quietly knows about you — screen, hardware, GPU, locale, network — none of which ever leaves your device. Every value labeled by source, computed live, and stored nowhere. Built to grow into a suite of small dev tools.

The one that started the spree — a real-time multiplayer, browser-based drafting game. Inspired by the dice-drafting of Sushi Roll, but very much its own thing: its own mechanics, a tapas-bar theme, and a healthy dose of take-that. 2–5 players with lobbies, built serverless so a pile of games can run at once for near-zero idle cost. Code-complete — getting a visual glow-up before the release candidate ships.

Not a garden gadget — it's a zone-based rotor-pressure estimator for irrigation field techs or your advanced hobbyist. You may be hamstrung by what you can measure at the spigot (static/dynamic pressure and available flow), but what a sprinkler actually delivers depends on the pressure way out at each rotor down the line. Every number is grounded in real hydraulics — Hazen-Williams friction loss, elevation head, orifice-law nozzle flow, even Joukowsky water-hammer — solved as a coupled lateral against published Hunter nozzle-performance tables, so nozzle selection stops being a guess. A pure, deterministic engine with a thin UI on top.

My own working process, codified as a Claude Code plugin anyone can install. It lifts the self-improving learnings loop out of claude-constitution — where each session records what it learned and self-corrects on the next run — and ports it into reusable slash commands, subagents, and hooks. (Aptly named — a palinode is a formal recantation, an author taking back a position they argued in an earlier work. That's the whole idea here: each run learns from the last and corrects course.)

Because "it feels better" isn't a metric. A harness that measures whether a change to an AI's system directives actually shifts its behavior: adversarial "trap" prompts run against each directive, structural rules auto-graded, the substance blind-judged by a separate model, and results aggregated with Wilson confidence intervals and a two-proportion significance test. Evidence over vibes.

My discipline-in-a-box — a reusable Claude Code “constitution” I stamp onto every new project. One CLAUDE.md, re-read in full at each session start so it survives compaction, enforcing the same non-negotiables everywhere: exhaustive spec before any code, strict TDD, independent audit & review subagents at phase boundaries, and a sign-off protocol for exceptions. Every project in this list is stamped from it.