Inspiration & References¶
The ideas, tools, and patterns that inspired the GYC Wiki architecture.
Primary Inspiration¶
Karpathy's LLM-Wiki Pattern¶
Source: gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
The core idea: instead of using RAG (retrieving from raw documents every time), the LLM incrementally builds and maintains a persistent wiki — a structured, interlinked collection of markdown files. When you add a new source, the LLM reads it, extracts key information, and integrates it into the existing wiki.
Key concepts we adopted:
- Three-layer architecture: Raw sources → Wiki → Schema
- Operations: Ingest, Query, Lint
- index.md — Content-oriented catalog of every page (the LLM reads this to navigate)
- log.md — Chronological record of operations (append-only, grep-parseable)
- "Compile, don't retrieve" — Knowledge is built up once and kept current, not re-derived on every query
- The human curates; the LLM does the bookkeeping — Writing, cross-referencing, and maintenance
Why it resonated with GYC: The pattern maps perfectly to a construction company's need — institutional knowledge that compounds over time, maintained without the burden that typically causes wikis to die.
My-Brain-Is-Full-Crew¶
Source: github.com/gnekt/My-Brain-Is-Full-Crew
A multi-agent system for managing an Obsidian vault using Claude Code. 8 AI agents + 13 specialized skills for knowledge management, email triage, and meeting transcription. Uses a PARA-inspired folder structure.
What we borrowed: - Hybrid PARA + domain-specific folder structure - Voice transcription → structured meeting notes pattern - Agent coordination model (dispatcher → skills → agents) - YAML frontmatter for metadata standardization
What we adapted: - Designed for a single person → we need multi-user company access - Personal wellness agents → replaced with construction-specific categories - Obsidian-only interface → we need web access via OPS Dashboard too
Related Tools & Projects¶
| Tool | What It Is | Relevance |
|---|---|---|
| Obsidian | Local-first markdown knowledge base | Primary wiki editor |
| WhisperX | Speech-to-text with speaker diarization | Voice transcription pipeline |
| Ollama | Local LLM hosting | Structuring transcripts, future AI queries |
| MkDocs Material | Beautiful static site from Markdown | Future read-only team access |
| qmd | Local markdown search engine | Future search when index.md isn't enough |
| Quicky Wiki | LLM-wiki implementation with confidence scoring | Reference for future knowledge metabolism features |
Key Principles (From Research)¶
- Format durability: Markdown is plain text (UTF-8). It will be readable on any computer for 30+ years. No vendor lock-in.
- AI-native format: Markdown is the native language of LLMs. Every page we write is instantly consumable by future AI agents.
- Git for provenance: Every change tracked by who, when, and why. Full audit trail.
- Compile once, query many: Build the knowledge structure incrementally rather than re-discovering it from raw documents every time.
- Humans curate, machines maintain: The team decides what knowledge matters. AI handles the cross-referencing, formatting, and maintenance.