Most AI work disappears at the end of the chat.
Prompts are not an operating system. They cannot reliably tell the next person what is true, what was approved, which source owns a changing fact, or what must not happen without permission.
The goal is not a giant second brain. It is making the right information easy to find, trustworthy enough to use, and structured so an agent knows its boundary.
An AI Brain makes useful work repeatable. It turns scattered knowledge, operating decisions, and workflows into a context layer that survives tools, devices, team changes, and time.
Markdown vault
The canonical, portable knowledge layer.
Obsidian
The human interface for writing, links, graph, Canvas, and search.
Obsidian Sync
Live transport for the same Markdown vault across enrolled devices.
Git + GitHub
Version history, review, recovery, and controlled replication.
Live systems
The source of truth for changing CRM, commerce, finance, campaign, and customer state.
Markdown vault → canonical durable knowledge Obsidian → human navigation and writing Obsidian Sync → live cross-device transport Git + private GitHub → history, review, recovery Live systems → current operational truth
Make the relationship between things explicit.
Interconnectedness is not linking everything to everything. It is creating a few reliable paths from source to fact to decision to action.
20-brands/A self-contained dossier: facts, offers, research, claims, strategy, voice, and brand-local skills.
30-projects/A continuation-ready control plane: current state, next actions, decisions, dependencies, and handoff.
50-operations/Cross-brand rules, policies, and repeatable operating procedures.
60-skills/Reusable agent-operable methods with prerequisites, outputs, and verification.
70-templates/Copyable structures—not the authority for a live decision.
One mutable fact should have one canonical owner.
Link to it. Do not copy it into five briefs and hope those copies remain true.
How a guest agent becomes useful without becoming dangerous.
Every serious vault needs a short root contract—often called AGENTS.md. It is tool-neutral rule of law: the same baseline for a teammate’s agent, your own local agent, or a new model next year.
- Read the root contract, then scoped project or brand context, before taking action.
- Treat facts, evidence, approvals, strategy, and instructions as separate objects.
- Use the relevant live system for changing operational truth; do not use a stale Markdown copy.
- Do not invent facts, sources, approvals, customer results, or claim permissions.
- Do not publish, spend, send, delete, or change a live system without explicit scoped authorization.
- Leave a verifiable handoff: source links, updated project state, a clean diff, and a coherent commit.
source artifact → verified fact → evidence / limitation → approval or decision → strategy → execution
Use both. Give each a different job.
Obsidian is the human interface.
Write in it. Follow backlinks. Search across the vault. Use Canvas when the work is spatial. It operates on the same Markdown that agents read.
Obsidian Sync moves current files.
Use it as the live transport layer across enrolled devices. Avoid editing the same file on two devices before sync is current.
Git and GitHub explain change.
They provide reviewable diffs, recovery, controlled replication, and a durable history of decisions. They are not a competing live sync engine.
Obsidian Sync moves files.
Git explains how they changed.
Do not run iCloud, Dropbox, Google Drive Desktop, Syncthing, or another unmanaged sync engine against the same vault. Multiple sync engines create conflicts that look like AI mistakes but are really transport mistakes.
Create a private Git repo and open it as an Obsidian vault.
Add AGENTS.md, CLAUDE.md, and INDEX.md at the root.
Create one brand/workstream dossier and one continuation-ready project.
Move one recurring workflow into a skill with a verification step.
An AI Brain should make agents reliably useful—not merely informed.
Find the right context. Know what is true. Know what is approved. Know what is unknown. Do useful work without creating hidden risk. Leave the system clearer than you found it.
Talk to Bryan about building this →