Personal Assistant Kit

An open-source system that turns an AI assistant into a personal chief of staff — one that reads your notes, triages your email, processes your meetings, and manages your task system. You stay in control; it handles the admin work.

Built on the Model Context Protocol (MCP), so it works with any MCP-capable assistant: Claude Desktop, Claude Code, Kiro, or whatever comes next. The meeting pipeline's background automation runs on Claude Code.

What this actually saves you

The daily grind that eats your morning:

• Sorting through inbox captures and deciding what goes where
• Reading email, flagging the 3 that matter, drafting replies
• Checking your calendar and figuring out what to prep for
• Writing status updates for your PM tool

The weekly maintenance that slips when you're busy:

• Reviewing your full task system for stale items and missing next actions
• Catching commitments you made in meetings that never became tasks
• Auditing waiting-for items that haven't moved

The meeting overhead that compounds:

• Turning a 30-minute conversation into structured notes, tasks, and follow-ups
• Remembering who said what, who committed to what, and what the unresolved tensions were
• Building a living profile of how each person communicates, decides, and collaborates

All of this happens through your existing tools (Obsidian, email, calendar, Slack) with a single rule: the assistant proposes, you approve. Nothing sends, posts, or deletes without your explicit say-so.

Who this is for

You're a good fit if you:

• Spend more time managing work than doing work
• Have 5+ meetings a day and can't remember Tuesday by Thursday
• Use (or want to use) a local notes app like Obsidian as your productivity system
• Want AI to handle the organizational scaffolding, not make your decisions

You don't need to be technical. The setup guides assume no coding experience. You do need about 30 minutes for initial setup and a willingness to keep your notes in markdown files.

What's in the kit

Core (works out of the box after setup)

Advanced (patterns for power users)

How it works

┌─────────────────────────────────────────────────────────────┐
│                     Your AI Assistant                       │
│         (Claude Desktop, Claude Code, Kiro, any client)     │
├─────────────────────────────────────────────────────────────┤
│                   Model Context Protocol                    │
├──────────┬──────────┬──────────┬──────────┬─────────────────┤
│ Obsidian │  Email   │ Calendar │  Slack   │  PM Tool / etc  │
│  (vault) │          │          │ (drafts) │                 │
└──────────┴──────────┴──────────┴──────────┴─────────────────┘

MCP is the layer that gives your assistant hands. Instead of just generating text, it can read your notes, check your inbox, look at your calendar, and write back — all through a standardized protocol.

The system prompt tells the assistant how to think about your system: what the folder structure means, how to process inbox items, how to run a weekly review, and what your communication preferences are.

Background automation (optional) keeps the meeting pipeline running unattended. On Claude Code, a bee-processor subagent drains new captures into notes, driven by the /process-bee-inbox command or a scheduled runner; PowerShell scheduled tasks handle the Bee sync itself. (The older Kiro hooks that once did this are retired; the weekly-status-update Kiro hook is unrelated and still available.)

The human-in-the-loop rule means the assistant gathers context freely (reads are auto-approved) but always asks before taking action that's visible to others (sends, posts, deletes require your confirmation).

Using it day to day

Once you're set up, the first thing to read is docs/USAGE.md: a day-in-the-life walkthrough, a copy-paste prompt library, the top workflows end to end, and what to expect. It answers the question setup guides never do, which is what do I actually type?

A few to get you started:

start my day

Runs your full morning triage (email, calendar, chat) into one dated plan.

process my inbox

Routes each captured item to the right GTD folder, on your confirmation.

run my weekly review

Flags stale items, missing next actions, and commitments that never became tasks.

what do you know about my system and how I work?

A good first prompt after setup: confirms the assistant read your configuration.

Full library and worked use cases: docs/USAGE.md.

The methodology

The kit is structured around Getting Things Done (GTD) — a system for managing commitments without keeping them in your head. You don't need to have read the book. Here's what matters:

1. Capture — Everything goes into an inbox. Don't organize it yet.
2. Clarify — For each item: is it actionable? What's the next physical step?
3. Organize — Put it where it belongs based on what it is and when you need it.
4. Reflect — Review the system regularly so you trust it.
5. Engage — Work from the system, not from memory.

The assistant handles steps 2-4 for you. You capture (step 1) and decide what to work on (step 5). The vault structure mirrors these stages:

00 Inbox/                  ← Everything enters here
01 Next Actions/           ← Things to do (Deep Work / Quick Wins)
02 Personal Projects/      ← Multi-step efforts with a defined outcome
03 Family & Personal Planning/  ← Household logistics, family coordination
04 Someday Maybe/          ← Not now, but not forgotten
05 Reference/              ← Keep but don't act on
06 Waiting For/            ← Blocked on someone else
People/                    ← Living bios of people you work with

Getting started

What you need

Setup (about 30 minutes)

Quickest path — run the setup wizard:

node scripts/setup.mjs

It asks your name, vault path, and which AI client you use, then generates all config files. Or follow the manual guides:

• Mac → SETUP-MAC.md
• Windows → SETUP-WINDOWS.md

Problems? → TROUBLESHOOTING.md

Optional add-ons

Example: what a processed meeting looks like

When a Bee recording finishes, the pipeline produces three outputs:

Tasks (00 Inbox/Bee/2026-05-07_weekly-sync_tasks.md):

- [ ] **Complete audit of expanded-scope products** — scope resourcing needs before May 27 meeting
- [ ] **Implement time-boxing defaults in Close the Loop** — 30-60 day rolling window; gate leadership rollout on this
- [ ] **Follow up with Jim on beta testing** — integrate Common Room data source this month

Meeting notes (05 Reference/Meeting Notes/2026-05-07_weekly-sync.md):

# Weekly Team Sync — 2026-05-07

## Topic Summary
Weekly sync covering team transitions, engagement strategy, and infrastructure planning...

## Key Decisions
- Stakeholder roadshow targeted for June
- Data consistency concerns gate leadership rollout of new tool

## Things to Keep in Mind
- New engineers took over from Simon; latency issues resolved
- Promotion review deadline May 21 for 6-7 cases

People notes (People/John Smith.md):

## Communication Style
Warm and personable in 1:1 settings. Opens with genuine personal connection before work.
Receptive to boundaries — doesn't push when told "not yet."

## Decision-Making Pattern
Seeks alignment before advocacy. Integrates rather than competes.

All of this is generated automatically, with privacy redaction applied. Personal/medical/family content is excluded. You review and route the tasks like any other inbox item.

Customizing

The system prompt is designed to be forked. Key files to make yours:

See docs/customizing-your-system-prompt.md for detailed guidance.

Built by

Joe Garvey — Head of Developer Technology

Other projects in this stack:

• SpendSense — Self-hosted personal finance automation

License

MIT — fork it, make it yours, share what you build.