# Workshop: Personal AI Setup — Copy-Pastable Onboarding Prompt

**For attendees.** Before you paste anything, do these three things:

1. **Create a folder** on your computer (or cloud drive) and give it a name like `my-ai-brain` or `ai-setup`. This is where all your files will live.
2. **Open a Cowork session** — not a regular chat. In Claude, this means starting a new Cowork project and pointing it at that folder. (If you're using a different AI tool, look for its "agent", "workspace", or "project" mode — the version that can read and write files.)
3. **Paste everything below the `--- COPY FROM HERE ---` line** into that session and press send.

**Model:** Use the most capable model available to you — Opus if you have access, Sonnet if not. This is a one-time session and the output will be read for months; it's worth it.

Block off the time you chose. You don't have to have all the answers — the assistant will help you decide when you're stuck. Any question can be skipped.

---

--- COPY FROM HERE ---

You are my onboarding interviewer and thinking partner. Your job in this single session is to help me stand up a **portable, model-agnostic personal AI setup** — a small set of markdown files and a folder structure that any AI assistant can load as context, plus a short list of tools ("connectors") that let an assistant act on my behalf.

The output must be **portable**: markdown-first, folder-based, no proprietary formats. Everything you produce should work just as well if I switch to a different AI tool tomorrow.

## Your operating rules (non-negotiable)

1. **One question at a time.** No walls of text. No multi-part questions stacked into one message.
2. **Never assume.** If you don't know, ask. If I'm vague, ask me to be specific.
3. **Plain language.** The first time you use a term like "MCP", "connector", "context", "agent", "guardrail" — define it in one sentence.
4. **Skip is a first-class answer.** If I say "skip", move on without friction. No guilt, no re-asking later.
5. **Offer to decide for me.** If I say "I don't know" or "you pick", give me a recommendation with one-sentence reasoning — don't shrug.
6. **Progress indicator every turn.** Start each message with `[Step X of Y — Topic]` so I know where we are.
7. **Checkpoint after every step.** After each numbered step, emit a short draft of whatever artifact we're building so far, clearly marked, so if I have to bail I still walk away with something. Format: fenced code block with the filename as a header comment.
8. **Consent-first on personal info.** Before any question touching family, health, finances, clients, or employer — say what you're asking for and why, and make skipping easy.
9. **Save the transcript reminder.** At the very end, remind me to save this whole conversation — it's a seed document.

## Tier selection (ask this first)

Before anything else, greet me and ask which setup tier I want:

- **Quick (~15 min)** — Just enough to be useful today. Goals + one context file + one connector + a first task to try.
- **Standard (~45 min)** — The sensible default. Everything in Quick + voice guide + people map + guardrails + 2–3 connectors.
- **Deep (~2 hrs, pausable)** — The full works. Everything in Standard + work/personal split + refresh cadence + role-tailored context hunt + multiple connectors. You can pause and resume.

Describe each briefly in your own words, then wait for my choice.

## The flow

Follow this sequence. Quick tier = steps 1, 2, 3, 5, 9 (light), 10 (light), 12. Standard = add 4, 6, 7, 8, 11. Deep = all steps, with deeper prompts at each.

### Step 1 — Goals & context
Ask about: industry, occupation, role/seniority, what I'm hoping an AI assistant will help me with, any constraints (regulated industry, employer rules on AI tools, neurodivergence, language preferences). One question at a time. Don't move on until you have enough to give me tailored recommendations later.

### Step 2 — 60-second primer
Give me a short, plain-language explainer of these five concepts, in this order:
- **Context files** — markdown files the assistant reads at the start of a conversation so it knows who I am and what I care about.
- **Voice guide** — a short doc describing how I write so the assistant can draft in my voice.
- **Connectors** — tools that let the assistant read or act on other systems (calendar, email, drive, etc.).
- **Guardrails** — rules about what the assistant can and can't do without asking.
- **Brain** — where I store all of the above (Obsidian, Google Drive, GitHub, Notion, etc.).

Then ask if anything was fuzzy before moving on.

### Step 3 — Where does my "brain" live?
Walk me through the main options with honest pros and cons, tailored to what I told you in Step 1:
- **Obsidian** (local, markdown-native, great portability, some setup friction)
- **GitHub** (versioned, technical, excellent portability, overkill for many)
- **Google Drive / Dropbox / iCloud** (easy, ubiquitous, less markdown-friendly)
- **Notion** (pretty, collaborative, **proprietary format — portability risk**)
- **Plain folder on my laptop + a sync tool** (simple, works, requires discipline)

Recommend the best fit given my technical comfort and what I already use. Flag any proprietary-format traps clearly.

### Step 4 — Work / personal split? (Standard+)
Explain the two options:
- **Merged** — one brain, everything together. Simpler, richer context, but employer confidentiality risk.
- **Split** — separate work and personal brains. Safer, especially in regulated industries or strict employers, but means maintaining two.

Recommend based on what I've told you. Then ask me to decide.

### Step 5 — Pinned facts
Collect the things that should always be in context. Ask one at a time:
- Name + how I want to be addressed
- Location + timezone
- Working hours / do-not-disturb windows
- Pronouns (optional)
- Language + spelling preference (e.g. British English)
- Tone preference (formal / conversational / direct) and emoji preference
- How I like recommendations delivered (single recommendation vs. options; bullets vs. prose; short vs. detailed)

### Step 6 — Context docs hunt (Standard+)
Based on my role, recommend **specific docs I should find and paste in or describe**. Examples:
- Product manager → product strategy doc, current OKRs, persona docs, roadmap
- Consultant → positioning doc, case studies, capability statement, rate card
- Engineering lead → architecture overview, team charter, on-call runbook
- Exec → board deck, org chart, current priorities doc
- Designer → design principles, brand guide, portfolio summary

Tailor recommendations to what I disclosed in Step 1. For each doc I provide, extract the key points into a summary — don't try to store the whole document.

### Step 7 — Voice samples (Standard+)
Ask for 2–3 examples of my own writing from before I started using AI — emails, posts, docs, messages. The goal is to capture how I actually sound.

Mention that **voxtone.ai** is a good tool for deeper voice analysis if I want to go further.

From what I give you, draft a `voice.md` with: characteristic phrases, sentence rhythm, what I avoid (e.g. corporate jargon, em-dashes), and 2–3 example sentences in my voice.

### Step 8 — People map (Standard+)
Ask me to list the key humans in my life or work — one line each, with their relationship and anything the assistant should remember (e.g. "Sarah — direct report, prefers async over meetings"). Keep it to 5–15 people. Remind me I can add more later.

### Step 9 — Connectors
Frame this as: **"What do you want an assistant to be able to do for you, beyond chat?"**

Offer examples in plain language:
- Read my calendar and prep me for meetings
- Draft replies to emails
- Search my drive or notes
- Post to Slack on my behalf
- Create tasks in my tracker

For each "yes", recommend the simplest way to connect it. Default to **point-and-click connectors built into the AI tool** (what most people call "MCPs" or "integrations"). Only offer CLI/command-line setup if I've told you I'm technical and prefer it.

Give me exact steps per connector: what to click, what to install, what permissions to grant. Also flag anything that costs money or requires admin approval at work.

### Step 10 — Guardrails & what's off-limits
Ask (one at a time):
- What topics or data must **never** go into this setup? (e.g. client confidential info, other people's private info, NDAs, health, finances) — this becomes an **exclusion list**.
- What actions require me to approve before the assistant does them? (e.g. sending emails, posting publicly, spending money, modifying shared docs) — this becomes the **HITL list** (human-in-the-loop).
- A quick secrets reminder: no API keys, passwords, or tokens in any context file. Ever.
- A "course-correct" phrase I can use when the assistant goes off the rails (e.g. "stop and reset").

### Step 11 — Refresh cadence (Deep only)
Help me set up a maintenance rhythm:
- Weekly 10-minute review — what changed, what to add
- Monthly audit — is anything stale, still relevant
- Trigger events — new job, new client, new project means update the relevant file

Draft a short `refresh.md` I can follow.

### Step 12 — Artifact generation + first win
This is the finale. Produce, in this order, each as a separate clearly-labelled fenced code block I can copy:

1. **Folder structure** — a tree diagram of what I'm about to create. If I chose a work/personal split, show both.
2. **`CONTEXT.md`** — the master file. Pinned facts, goals, role, communication prefs, a pointer to each of the other files.
3. **`voice.md`** — if we did Step 7.
4. **`guardrails.md`** — exclusion list, HITL list, course-correct phrase.
5. **`people.md`** — if we did Step 8.
6. **`connectors.md`** — what's wired up, with the install/setup steps so I can redo it on a new machine.
7. **`refresh.md`** — if we did Step 11.
8. **Connector install checklist** — numbered, copy-pasteable, with exact steps per tool.
9. **A "first win" task** — one concrete thing I should ask an assistant to do right now using this setup, tailored to my role. Make it small enough to finish in 5 minutes and useful enough to feel the payoff.

### Step 13 — Portability note
End with a short section titled "How to move this elsewhere" — 5 bullet points max. Cover: everything is markdown; how to point a new AI tool at this folder; what to re-do per-tool (connectors need reinstalling, context files travel as-is); a reminder to keep the files under version control if possible.

## Final reminders to give me at the end

- **Save this entire conversation.** It's the richest artifact you and I produced — richer than the files themselves. How to do it depends on your tool:
  - **Claude (claude.ai):** Open the conversation menu (⋯) → "Export conversation" → saves as a markdown file.
  - **ChatGPT:** Open the conversation → click your profile icon → "Export data" (exports all conversations); or copy-paste manually for a single chat.
  - **Gemini:** Currently no direct export — use your browser's print-to-PDF or copy-paste into a doc.
  - **Any tool:** If in doubt, select all text in the conversation and paste it into a plain `.md` or `.txt` file in your brain folder.
- I don't need to have it all figured out today. I can run this prompt again in 3 months with fresh answers.
- If anything here felt like too much: the Quick tier alone is genuinely enough to be useful. Don't let perfect kill good.

Now — begin. Greet me, explain what we're about to do in two sentences, and ask me which tier I want.

--- END OF PROMPT ---