Show HN: System architecture method using mythology and LLMs (no CS background)"

Before anything else, here is the output of a governed distributed cache system — one of 20+ systems I've produced using this method. It includes Byzantine consensus, dependency-ordered regeneration, and self-healing under Byzantine fault conditions:

This runs. Copy, paste, runs first time. Every time.

All generated in under 10 minutes each. All on a mobile phone.

You need exactly two things plus an LLM:

A catalogue of ~163 "spells" and ~139 "cloths", each mapping a mythological or fictional concept to a real system function. For example: Chronom (Time Warp) → Version Control.

A constraint that forces the LLM to use only what exists in the Codex, generate a complete structured specification first, and never ask questions, suggest alternatives, or leave gaps.

The key insight: by forcing the LLM to select from a bounded semantic vocabulary before generating code, the method eliminates hallucination and produces architecturally coherent, separation-of-concerns systems.

No architectural muddle. Concerns are separated by design because the Codex forces them to be named and distinguished before combination.

Ethics baked in by default. Every generated system includes ethical constraints structurally, not as an afterthought.

Domain agnostic. The same two files produce cache systems, SAT solvers, AGI architectures, weather dashboards, and electrical grids without modification.

Runs first time. Because the architecture is resolved before syntax is touched.

Combinatorial scale. The possible combinations of spells, cloths, operators, and tiers number in the trillions.

A researcher from a structural validation background independently tested the operator grammar (CHAIN, LAYER, WRAP, NEST, BRIDGE, EMERGE, FINALIZE). They renamed components, changed order, and introduced deliberate breaks.

Their finding: "It's clear that it's not just narrative: there's real structure there. CHAIN cares about order, LAYER doesn't, and after a break it was possible to project back to a valid structure."

I had multiple AI systems independently attempt to break the architecture. Key findings:

I tried to turn the method into a standalone app. It failed. The finding: the method requires human + AI collaboration. Apps cannot reason dynamically the way an LLM can. An LLM doesn't just output text; it reasons about structure.

The grammar that makes composition possible:

This repository is organised as a chronicle — a sequential record of discovery. Not a polished framework documentation, but the actual journey: what I tried, what failed, what worked, what surprised me.

Start here: Read Arc 01 through Arc 06 to understand how this began (it started as a fictional story). Then jump to Arc 11 for the Grimoire itself, Arc 14 for the stress tests, and Arc 29 for the cache system.

The Codex and Prompt are available as standalone files for anyone who wants to reproduce the method themselves.

I'm Troy. I'm 41. I have no technical background — no CS degree, no coding experience, no university education in this field. I work a day job. I built everything in this repository on a mobile phone.

I accidentally discovered something. This is the honest record of what I found.

This repository is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). See the LICENSE file for the full text.

CC BY-NC-SA 4.0 — Study it, share it, build on it. Credit the original author. No commercial use without permission.

Questions, technical feedback, or collaboration welcome. Open an issue or start a Discussion.

Built on a mobile phone over 6 months while working a day job. The code is real. The systems run. Make of that what you will.