Lachlan Brown

Spirescape mixes Slay the Spire's card combat with RuneScape's open-world skilling and economy. I designed it before I built it. Eighteen rounds of a structured design interview produced 168 locked decisions, each with an ID, a date and the reason behind it, in one log that is the only source of truth. Superseded decisions get struck through, not deleted. Open questions have their own file. No document and no agent is allowed to invent an answer. Every rule in the code cites the decision it comes from, and because "everything connects to something" is the one design pillar, the content loader refuses to start if any material has fewer than two uses.

The prototype is a zero-art tutorial you can play in a browser. One shared rules library is compiled by both a headless .NET server and a Unity 6 client, so the client never works out a rule itself. It sends what you meant to do and draws whatever the server's 600 ms tick sends back, and a second browser is a second player in the same world. The test suite plays the whole tutorial end to end over a real socket with a headless client, and a smoke driver plays the opening in a built player and saves screenshots. It runs on a five-dollar Lightsail box in Sydney behind Caddy, which serves the WebGL build, fetches its own certificate and proxies the WebSocket on the same origin. The music is written as MIDI in Python and rendered through a General MIDI bank, the way the 2004 RuneScape soundtrack was.

MusicLearn sits on a MIDI track in Ableton Live, watches the keyboard, draws what I played on a real grand staff in the Bravura font, and checks it against the exercise. MIDI passes straight through, so instruments and recording keep working. It is C++ on JUCE 8, and the whole interface is one HTML file hosted in WebView2 and baked in at build time. Serve that file from a folder and it runs in Chrome against Web MIDI, so I can work on the interface without rebuilding the plugin. The drills and item pools are plain data in the same file.

The drills are built around recall, not copying. A new item is shown once with its answer, then hidden. A wrong note is drawn in red exactly where it sits, next to the note that was asked, with a message like "F5, one step too high". Misses come back within a few items, the picker weights everything by error rate and how long since you last saw it, and a level suggests moving on at 90 percent first-try over the last twenty. Progress lives in the plugin's WebView2 profile. Nothing leaves the machine.

I do not model anything by hand. Each character and tower is a Python script that runs Blender headlessly and exports an FBX. The script's knobs sit at the top as plain variables, so trying a design change means writing a row of overrides, not editing code. A preview harness renders each row in parallel Blender processes, stops them fighting over the GPU, and writes every wave to its own folder with a manifest and the renderer version, so I can reproduce any picture later. Previews cannot touch the real .blend or FBX. A hook interlock refuses to run any script that could.

The agent workflow around that harness is on its third design, and I rebuilt each one on what the last one measured. The first was one agent with one fixed camera, editing the script against its own render. The second was a fleet: an orchestrator, designers, modellers working in their own git worktrees, and three verifiers scoring renders out of 100. Across 81 logged attempts the panel almost never agreed, so I turned every rule that had been prose into a mechanism. The third dropped the scores for blind pairwise judging on a random 360-degree orbit, added silhouette masks and sanity checks, and runs per-asset shepherds and makers through review panels that can accept, send back or scrap. I steer it from a live page with a chat box. A champion is a run, a cell and a seed, so I can replay any verdict.

Earshot transcribes whatever it hears in real time, shows the words in a floating overlay as they are spoken, and fires rules I have written on what was said. Silero voice-activity detection gates a two-tier Whisper pipeline: a tiny model gives rough partials every 400 ms while you are still talking, then an accurate model transcribes the finished sentence. Rules only ever see the final text, so nothing fires on a guess.

Rules escalate through three tiers: keyword and regex for free, a local model through Ollama, then a budgeted cloud call. Actions go out over MQTT, HTTP, the shell and AutoHotkey. Nothing leaves the machine unless a rule says so, and the only thing written to disk is the list of triggers that fired.

Lodestone backs up, restores and shares a Claude Code and terminal setup across machines. Copy a config folder to a new computer and half of it is dead on arrival: a hook pointing at a file that got left behind, a script with no runtime, a server with no key. Lodestone treats each piece as a capability with dependencies, checks the destination is ready before restoring, and installs whatever is missing.

It finds secrets across every provider and never stores them in plaintext. A desktop app and a standalone CLI share one Rust engine, split across core, platform, CLI and Tauri crates.

A set of dashboards that render inside Obsidian, each one a single JavaScript block over a Python backend. The mail cockpit merges several Gmail accounts into one feed with triage, compose, reply and full Gmail search, and sorts every conversation into Work, Security, Finance, Releases, Marketing or Personal. Rules go first and a small model handles the leftovers. Appointments it finds in mail land on their own calendar automatically.

Calendar and tasks sync both ways with Google. A process warden keeps the long-running services alive and reports on them. OAuth tokens are DPAPI-encrypted and kept outside the vault.

The Sutherland Shire's water asks whether ANSTO and the Lucas Heights tip can reach the Shire's drinking water (they can't: both drain downstream of Woronora Dam), what coal mining under the catchment is doing to it, what the PFAS record from 2024 and 2025 actually says, and how a resident could test for themselves. It is built from licences, monitoring data, regulator reports and inquiry transcripts, with 87 linked sources and no water sampled.

The State Harm Index scores the governments of 37 countries from 1 to 100 on repression, violence, environmental damage, corruption and social harm, with the change over one day to ten years. Yearly indicators from V-Dem, Transparency International, UCDP and Our World in Data set the long run; the day's news, scored by a small model, moves it. It is built to rebuild itself every evening; that run is paused for now, so the page shows the data to 7 October 2026.