How I Do It
Example setup

An illustrative ecosystem, separate from your learning workspace.

Explore the example · Adapt this setup

Steal this

Copy the practice, not the life

You do not need Plaud, Grok, and five models on day one. You need a brain (git), a printed chart (Drive or paper), one loop, and named gates. Wispr Flow, Speechify, WhatsApp, Inkbox, Cursor, Netlify, and a GoDaddy domain can wait until the morning loop is boring. When you do add jobs, put them on the roster — one box per schedule, tools split, never treating “separate bots” as a firewall.

  1. 1

    Make a brain

    Create a private GitHub repo. Add AGENTS.md (how agents should read it), a charter/values file, and one instructions file for your daily bot. Markdown only.

  2. 2

    Make a human layer

    Export selected GitHub Brain documents into stable Drive documents for reading. Preserve sharing and stamp the source commit; do not maintain a competing source.

  3. 3

    Plug the right sockets

    MCP for tools that publish a server (voice, tasks). Direct OAuth for Workspace if your model offers it. Connectors/plugins for everything else. Git push for code (Cursor → GitHub). DNS for the public name (GoDaddy → Netlify). Wispr Flow for dictation, Speechify for listening, Plaud for recordings — three different microphones. Do not use git to store mail or WhatsApp.

  4. 4

    Write one loop

    Start from the canonical morning routine. Keep staging and live Todoist task state. Test one loop before adding more.

  5. 5

    Draw the wires

    When you add a tool, add a node and an edge whose type matches the data. If you cannot name the data, you are not ready to connect it.

  6. 6

    Name the gates

    Sending, paying, deleting, publishing, submitting, DNS: human. Staging promotions: human. Research and filing: agent.

  7. 7

    Put a name on the public copy

    When you are ready for a real URL: Netlify hosts the scrubbed site, GoDaddy points the domain. DNS is a human gate. Cursor is how the site is edited — push to GitHub, never treat the IDE as a second brain.

The one prompt to start

Paste into a Grok Bot, a Claude Project, a Custom GPT, or a Gemini Gem. Change product names. Keep and the .

You are my weekday morning coach. Be warm, specific, and brief.

1. Read shared instructions from my private GitHub repo first.
2. Read this week’s Google Doc and the latest dated notes in Drive.
3. Pull new voice recordings (last 24h). Extract actions as STAGING tasks — never onto the main list.
4. Read today’s calendar. If the first event is earlier than 7:30, start 10 minutes before it.
5. Triage with me: staging, weekly goals, free blocks. I decide what is promoted.
6. Create calendar events whose titles are the actual task.
7. Append a short dated note so tomorrow has memory.

Never send email. Never promote staging without my say-so.

Then add loops in this order

  1. 1Weekday morning planTue–Fri, 7:30 AM (or 10 minutes before the first block)
  2. 2Evening done listWeekdays, 5:00 PM
  3. 3Friday reflectionFridays, 4:00 PM
  4. 4Monday weekly goalsMondays, 7:30 AM
  5. 590-day goal cascadeTuesdays, 9:00 PM
  6. 6GitHub → Drive instruction syncWhen instructions change
  7. 7Inkbox secretaryWhen mail arrives from @inkbox.ai
  8. 8Speak and listenWhenever typing would slow you down
  9. 9Ship the public playbookOn merge to main · DNS only when the name changes

Words used on purpose

MCP
Model Context Protocol. An open standard (from Anthropic) for connecting an AI to tools and data. USB-C for AI.
Direct OAuth
The AI product logs into the app as you, with no plugin broker. ChatGPT → Gmail is direct.
Connector
A native plugin, or — only as gap-fill — a broker such as Composio. Native Grok Bot plugins first (`std-grokbot-connectors`).
Agent
A model that decides its own next step and may use tools until the job is done.
Workflow
A model plus tools on a path you defined in advance. Predictable. Safer for mail and money.
Orchestrator
The agent that breaks work up, assigns it, and synthesizes — it should not do the work itself.
Staging
A waiting room for tasks. Recordings land here. You promote them. Nothing jumps onto the main list.
Human gate
A step only a person can pass: send, pay, delete, publish, submit, promote.
Git brain
GitHub Content holds playbooks and docs (how-i-do-it, governance markdown). GitHub Code holds application repos. Diffs are the audit log. Survives switching models.
Drive layer
Google Docs humans actually read. A stamped copy of the brain, not a second brain.
ReAct
The loop almost every agent runs: Perceive → Reason → Act → Reflect.
A2A
Agent-to-Agent protocol. How two agents from different vendors talk. Complementary to MCP, not a rival.
Least privilege
Each agent gets only the tools its job requires. The secretary does not get Plaud.
Charter
A Markdown file of your values. Coaching agents read it and are forbidden to substitute their own.
OWASP
Open Web Application Security Project. Their 2026 Top 10 for agentic apps is the checklist this security page follows.
Council
Several role-named agents each review the same input; a synthesizer writes one memo. Disagreement is useful when they share a charter.
Wispr Flow
OS-level voice dictation, plus a remote MCP (OAuth, read-only) for Notetaker meetings, transcripts, scratchpad notes, and share links. Dictations are not on that wire. Hands, not a recorder — Plaud remains the meeting record.
Speechify
Text-to-speech. The chart, read aloud. Ears, not a writer.
Inkbox
Identity and communication layer for AI agents: email, phone, iMessage, internet address, Agent-to-Agent. One badge per role.
WhatsApp
Fast human messaging. Same gates as mail: draft, then a person sends.
Grok Bot
A persistent scheduled agent from xAI. Many schedules can share one bot. They share infrastructure, so they are not a security boundary — split tools and prompts instead.
Team
Named specialists around one record. They coordinate through git, not private chat. Cadence, Grant, Sherpa, Values.
Roster
The list of actual jobs — Monday coach, Inkbox secretary, Scout, Counsel — as opposed to the AI tools. Named agents and their teams appear in Team structure.
Cursor
The local workbench. AI can edit inside it (Composer, Claude Code). Nothing is shared with the other AI tools until you push to GitHub.
GoDaddy
The registrar. It holds the public name of the playbook site. Netlify is the building; GoDaddy is the street address.
Governance
A small private repo of principles, standards, and guides that every agent reads before it works. Projects cite document ids instead of copying text.
Causal chain
The commit message shape from the governance repo: GOAL, OBSTACLE, DECISION, PRINCIPLE, CONSEQUENCE. The diff shows what changed; the message carries why.
Claude Code
Anthropic’s terminal agent. Runs on the local drive inside one project folder, reads and edits files, runs commands, and pushes to GitHub when asked.
RLS
Row-level security. Rules inside Postgres (Supabase) that decide, per row, who may read or write. The layer that actually protects data when the browser holds a public key.
Cloudflare
Live DNS, caching, and bot filtering in front of the site. GoDaddy keeps the name; Cloudflare answers the lookup. The Grok Bot plugin is authorized. Moving nameservers is still a publish.
DNS
Domain Name System. Records that point a name (the website people type) through Cloudflare at a host (Netlify). Changing DNS is a publish.
Hugging Face
Model and paper hub. Grok Bot reaches it over MCP; Cursor reaches it over a connector. Models, papers, GGUFs. No secrets in Hub uploads.
Trip Waffle
Live travel itinerary app. Remote MCP at tripwaffle.com/mcp. Grok, Grok Bot, and Claude each hold their own client — same multi-client pattern as Wispr. Not a second calendar; Google Calendar remains the calendar of record.
Composio
Gap-fill plugin broker only — Harvard Canvas, Substack, WhatsApp Cloud API — when no native Grok Bot plugin exists. Native plugins first (`std-grokbot-connectors`).