PHASE 0 · GROUNDWORK
🧠 3 Ms
🗺️ 7 Buckets
Mindset + scope — done before you build the stack
▼ then assemble the stack ▼
FOUNDATION
🗂️ The Repo
Synced, portable, tool-agnostic
LAYER 1
📚 Context
CLAUDE.md + context files — the brain
LAYER 2
🔌 Connections
API / CLI / MCP to real data
LAYER 3
⚙️ Capabilities
Skills — reusable SOPs
LAYER 4
🔁 Cadence
Routines run it while you sleep
INTERFACE
👤 You + your team
Talk to it; it does the work
▲ watch the stack build as you scroll
Before you build · the framework
The 3 Ms
How to think about AI before touching a tool — mindset first, build last.
- Mindset — change how you approach work. Three habits: default shift ("how could AI do ≥30% of this?"), function breakdown (your role is a tree of tasks — automate one reusable chunk at a time), curiosity rule (treat AI as a mentor, not a vending machine — ask why, ask what breaks).
- Method — decide what's worth automating and how much. Not everything should be; find each task's leverage %.
- Machine — the actual technical build (the AIOS itself). That's the rest of this map.
Why it matters: adoption is where AI projects die. Mindset gets it used, Method aims the effort, Machine is what you ship.
Before you build · scope
Map the 7 buckets
Sketch the whole business before wiring anything. This map becomes your build plan.
- Revenue · Customer · Calendar · Comms · Tasks · Meetings · Knowledge
- For each, list the real tools (Stripe, Gmail, ClickUp, Fireflies, Drive…).
- Every tool connects somehow — API, CLI, or browser automation.
Why it matters: this is the discovery worksheet on a client call — it drives the scope and the quote.
Foundation
The repo
Put the system on a private GitHub repo. Everything else sits on top of this.
- Syncs across machines; portable across tools (Claude Code, Codex, …).
- It's what cloud routines clone on every run — the substrate for autonomy later.
Why it matters: version-controlled + portable = a professional, recoverable deployment. Also the prerequisite for "while you sleep."
Layer 1 · Context
The brain
What the AI knows about you, the business, the voice, the money.
- Run the onboard skill — a 7-question interview that scaffolds context.
- Paste writing samples verbatim so it learns your real voice.
- CLAUDE.md = the master prompt. It evolves daily — never set in stone.
Why it matters: context quality decides everything downstream. Thin context = a generic AIOS nobody trusts.
Layer 2 · Connections
Reach the real data
Claude alone can only search the web. Wire it to where the business actually lives.
- API endpoints > MCP servers — MCP loads everything and burns tokens.
- Save a local markdown reference of endpoints once, reuse it forever.
- Secrets in .env, never chat. Give it a scoped account, not your keys.
Why it matters: token efficiency = lower run cost = better margins. Scoped keys = the safety story clients need.
Layer 3 · Capabilities
Skills — SOPs for the agent
Reusable recipes the AI runs consistently. The "what it actually does" layer.
- 6-step build: name/trigger → goal → meat → references → guardrails → improve loop.
- Built with the AI by doing the task, then "turn this into a skill." Expect 5–10 iterations.
- Progressive loading keeps dozens of skills cheap (name → full skill → references).
Why it matters: one good skill multiplies across the whole team. That's the scaling/ROI story.
The hard rule
Order is law
You cannot have Capabilities without Context. You cannot have Cadence without Connections. Build 1 → 2 → 3 → 4 — every time. Skip a layer and the one above it collapses.
Layer 4 · Cadence
It runs itself
Turn a skill into a schedule. This is what makes it an operating system, not a tool.
- Routines (cloud) — run with the laptop closed; "while you sleep." Stateless, repo + API only.
- Loops (local) — in-session, 3-day expiry; "help me this sprint."
- Ask: do I need help now, or every week forever? That picks the tool.
Why it matters: autonomous overnight operation is the headline promise to clients.
Interface · the daily reality
You live in it
Do as much as possible inside one window. Stop context-switching across 40 tabs.
- Daily: "help me plan my day." Weak plan? Note the missing context, patch it.
- Weekly: run audit + level-up. Push heavy-use skills to plain automations.
- Boring beats clever — deterministic workflows beat agents 9/10.
Why it matters: you can't scale (or sell) a system you haven't lived in. Personal AIOS first.