I build prompt systems, knowledge pipelines, public wiki workflows, and agent teams that run in production.
프롬프트 시스템, 지식관리 파이프라인, 공개 위키 자동화, 그리고 운영되는 에이전트 팀을 만듭니다.
It started with prompt engineering — treating prompts as a real engineering artifact, iterating one workflow through dozens of revisions until it consistently produced usable output across models. That discipline generated volume: prompts, research notes, drafts — thousands of markdown files that a folder tree couldn't hold anymore. So I built a knowledge pipeline — extract, structure, store, link — and open-sourced it as knowledge-manager; my Obsidian vault grew past 13,000 interlinked notes.
But a private knowledge base only compounds if it's used, so the next step was publishing: an automated wiki (akwiki.org) where community conversations become fact-checked articles, searchable archives, and keyword network maps — refreshed daily without manual editing. Running that publishing loop every day is more work than one person or one AI session can do, which is where agent teams came in: a team of specialized bots — research, scheduling, writing, engineering — each anchored to its own folder of the vault, coordinating over Discord, running around the clock.
Each layer exists because the previous one overflowed — prompts created knowledge, knowledge demanded structure, structure was worth publishing, and publishing required a team. The whole stack has been in continuous production for over six months. Along the way it turned into teachable, installable material: two distribution bundles (ThisCode, ThisCodex) and a hands-on Claude Code/Codex course built for a real lecture series. This profile is the logbook of that path.
시작은 프롬프트 엔지니어링이었다 — 프롬프트를 제대로 된 공학 산출물로 다루면서 하나의 워크플로우를 수십 번 개정했다. 그 습관이 수천 개의 마크다운 노트를 만들어냈고, 폴더 정리로는 감당이 안 돼 추출→구조화→저장→연결 지식관리 파이프라인(knowledge-manager, 오픈소스)을 만들었다 — 옵시디언 vault는 13,000개 이상의 연결된 노트로 자랐다. 쌓인 지식은 써먹어야 복리가 되기에 다음은 발행이었다 — 커뮤니티 대화가 매일 자동으로 검증된 기사·아카이브·키워드 지도가 되는 위키(akwiki.org). 이 발행 루프를 매일 돌리는 건 한 사람·한 세션으로는 불가능해서, vault 폴더 하나씩을 작업 공간으로 삼은 전문 봇 팀(리서치·일정·글쓰기·엔지니어링)이 Discord로 협업하며 24시간 돌아가게 했다. 앞 단계가 넘쳐서 다음 단계가 필요해진 경로이고, 전체 스택은 6개월 이상 실운영 중이다. 그 과정 자체가 가르칠 수 있는 자산이 되어 배포판 2종(ThisCode·ThisCodex)과 실강의(Claude Code/Codex 과정)로 이어졌다. 이 프로필은 그 길의 기록이다.
The publishing loop runs on a team of specialized bots, each anchored to its own folder of the vault and coordinating over Discord — around the clock. 발행 루프는 vault 폴더 하나씩을 맡은 전문 봇 팀이 Discord로 협업하며 24시간 돌립니다.
| Bot | Role |
|---|---|
| 🎛️ Karpathy | Orchestration & policy — coordinates the team and holds the workflow gates |
| 🔍 Konan | Research & fact-checking — cross-verified sourcing, contradiction detection |
| 📅 Strange | Schedule & tasks — calendar, todos, watchdog liveness |
| ✍️ Geuljaegyeong | Writing — Threads, essays, long-form drafts |
| 🎨 Sonseokhee | Image generation & code build — Codex-side engineering |
| 🐣 AK-Tofu | Daily community-insight pipeline — collect → summarize → publish |
| Project | What it is |
|---|---|
| tofable | Transfers a strong model's working style into a harness: rules, verification gates, and benchmarks. |
| knowledge-manager | Extract, structure, store, and link web/PDF/social content into an Obsidian or Notion knowledge base. |
| ThisCode / ThisCodex | Distribution bundles for Claude Code, Codex, Discord bot operations, and rules-system patterns. |
| prompt-engineering-skills | Prompt engineering skill library for model-specific workflows and image/video generation patterns. |
| lesson-cc-codex | Interactive Claude Code + Codex lesson engine for a hands-on lecture series. |
- Prompt engineering: prompts treated as versioned engineering artifacts.
- Knowledge management: Obsidian-centered extract, structure, store, and link pipeline.
- Publishing: akwiki.org, daily community articles, searchable archives, and network maps.
- Agent teams: specialized Discord-coordinated bots for research, schedule, writing, and engineering work.
- Teaching: installable bundles and hands-on Claude Code/Codex course material.
