Your knowledge base, written by conversation.
A position-addressed, harness-agnostic vault, scaffolded in one command and shaped by an interview that compiles into your purpose, identity, and workflow.
Node 20+. Click the terminal to copy.
The skill is a single npx command. It writes the position-addressed scaffold and registers itself with whichever LLM tool you already use.
Five systems, one tradeoff each. The comparison is the brand argument.
| System | The tradeoff |
|---|---|
| SystemKarpathy raw to wiki | TradeoffNo template. You build the structure manually. |
| SystemAI Impact infinite_brain | Tradeoff16 node types, 10 edge types. Over-engineered for solo operators. |
| SystemTiago Forte PARA | TradeoffDesigned for human re-reading, not agent retrieval. |
| SystemCole Medin AI second brain | TradeoffStatic identity files. No ingestion model. Claude-only. |
| SystemNick Milo Ideaverse-OS | TradeoffVault template only. No LLM build flow. |
| Systemideaverse-os | What shipsPosition-addressed scaffold + LLM-driven interviewer + harness-agnostic ingestion. |
What we kept
Raw to wiki rhythm. Ingest cheap, synthesize deliberately.
see: his recurring thread on personal wikis as memory
Position-addressed prefixes (00, 10, 99) so the LLM never has to guess where anything lives.
see: Linking Your Thinking, Ideaverse-OS template
Domain separation: work, personal, archive as first-class slots.
see: PARA Method, Building a Second Brain
Identity-as-files: who-you-are encoded in markdown the agent reads.
see: AI Agent Stack templates on GitHub
What we changed. A 5-phase interview compiles your purpose, identity, and workflow into the lean files. Every harness loads them through its own convention file. Capture is one turn, not a multi-step workflow.
npx ideaverse-os init scaffolds a position-addressed vault. Numbered prefixes mean the LLM never has to guess where anything lives.
soul, me, user, compass, memory, runbook
synthesis surface (your wiki articles)
first user domain, named in the interview
shaped by yousecond user domain, named in the interview
shaped by youingestion: web-clips, papers, youtube
spikes, prompts, customization
ideaverse-capture, cortex-*, web-clip-report
daily notes
attachments, screenshots, diagrams
retired sources and notes
Every query lands at 00-agentic-OS first — soul, identity, conventions. Then 10-cortex routes by position into a domain (20- or 30-). 40-raw feeds back upward via wiki-compile. Positions are addresses; no guessing.
Then /ideaverse-os build opens a 5-phase interview. Each phase asks a few questions and compiles your answers into one of the lean files. Skip nothing, re-run any phase later.
What you are optimizing for. This week, this month, this quarter.
Q. For each of those priorities, what's the success metric? How will you know you succeeded vs. just stayed busy?
Q. When two priorities collide, say a work emergency and a personal goal, which wins? Is there a rule, or does it go case-by-case?
## What I'm Optimizing For ### This week - Top bet: ship the build interview - Success metric: 1 user runs the 5 phases end-to-end ### When priorities collide work-emergency wins, log the swap in 70-daily/.
Who you are when you sit down to work.
Q. In 2-3 sentences: what drives you? What are you building toward over the next 1-2 years?
Q. What are you known for among people who know you well? What do they come to you for?
## What drives me Building tools that compound my own work and other people's. The agentic-OS pattern. ## Known for Receipts-first. Honest pushback. Bridging embedded depth with AI-tooling depth.
How you actually work, day to day.
Q. Which LLM harnesses do you use? Claude Code, Cursor, Codex CLI, Gemini CLI, ChatGPT web, Claude.ai web, Copilot CLI, other. For each, when do you reach for it?
Q. Which is your primary, the one you spend the most time in?
## Harnesses Primary: Claude Code (Sonnet 4.6 for build, Opus 4.7 for design and planning). Secondary: Cursor for refactors, Codex CLI for one-shots. ## Daily rhythm 06:30 /work-daily-note -> deep work 07:00-12:00
How your life splits across folders 20- and 30-.
Q. Which template fits your mental model? Work plus personal, creative, researcher plus builder, trader, day-job plus side-project, solo, or custom?
Q. Default name for your second domain is the template default. Use that, or rename?
template: work-personal my-vault/ +-- 20-work/ day job: AutoPath, Mytra +-- 30-personal/ career, portfolio, learning
Hooks, GitHub backup, WSL paths. All opt-in.
Q. You'll get three hooks: session-start (loads compass on every new session), load-memory (injects the lean files), save-session (logs transcripts). Confirm all three, or pick a subset?
## Session starts | Trigger | Chain | |--------------|----------------------| | Mon-Fri AM | /work-daily-note | | Weekend PM | /personal-daily-note | | Sunday PM | /weekly-review |
Phase 4 of the build interview offers six presets plus Custom. Each preset names your 20- and 30- domains. Hover a preset to see how the vault tree changes.
Blank slate, build your own split.
for: anyone whose life does not map to the presets
The classic split.
for: engineer with a day job and weekend projects
Make-things-first.
for: artist, writer, or musician with life on the side
Reading meets shipping.
for: indie hacker who research-then-builds
Markets first, life second.
for: full-time trader, part-time everything else
Both fit-for-purpose.
for: PM with a serious side project
One life, one folder.
for: founder, freelancer, generalist
Why the architecture works the way it does. Each page covers one load-bearing idea.
Architecture
How every LLM tool loads the same identity files through its own convention.
Read -->
Commands
What each command does, what each phase compiles, and the lock-file flow that keeps both safe.
Read -->
Phase 5
Claude hooks, GitHub backup, and WSL paths. Three opt-ins, each independent.
Read -->
Article standard
The required fields, typed-edge sources, and decay rules cortex-lint enforces across 10-cortex/.
Read -->