Personal AI Chief of Staff — Cursor + $20/mo

A personal AI Chief of Staff built inside Cursor (or Claude Code) on a $20/mo subscription. Runs four loops on a schedule: (1) inbox triage — pulls Gmail + Telegram + Slack via MCP, classifies into Reply Now / Reply Later / Acknowledge / Ignore with one-line summaries; (2) schedule sweep — checks the next 7 days, flags conflicts, suggests defragmentation, drafts decline-templates for low-value meetings; (3) Telegram aggregation — collapses all unread group-chat noise into 3 themed bullets per chat per day; (4) Notion / vault sync — picks up any meeting notes touched since yesterday and pushes structured action-items into the task tracker. Single owner, single repo, no SaaS.

One C-level operator built this for themselves in Cursor for $20/mo. Real, working, single-C-level deployment. Reported effects: inbox processed by 09:30 most days, Telegram aggregation killed the "I'm afraid to open the chat" reflex, and the schedule-sweep loop alone saved an estimated 4–6 hours per week of meeting that would otherwise have happened. Anchor cost reference: $20/mo per seat — concrete enough to put on a CFO's desk.

1. Create a single private repo ~/cos/ with the following structure:

• cos/AGENTS.md (or a one-line CLAUDE.md that contains @AGENTS.md) — the master rules file (see prompt below).
• cos/inputs/ — empty folder where MCP scrapers will drop JSON every hour.
• cos/state/ — vault of recurring people, projects, "do not interrupt" rules, your own role and constraints.
• cos/loops/ — one folder per loop: inbox/, schedule/, telegram/, notion-sync/. Each has a loop.md describing inputs, decision rules, outputs.
• cos/output/ — where the loop writes its result (Markdown file per day per loop).
• cos/scripts/ — small MCP scrapers and the cron runner.

2. Wire 4 MCPs: Gmail (read-only is fine), Telegram (your own session), Google Calendar, Notion (or whatever your vault is). Each MCP gets a small Python wrapper in cos/scripts/ that on schedule writes a JSON snapshot to cos/inputs/<source>.json.
3. Write the master rules file. This is the chief-of-staff "constitution" — what to do, what not to do, who to interrupt, who not to.
4. Write the four loop prompts. Each loop's loop.md is a runnable instruction the agent executes when invoked.
5. Schedule via cron. Hourly: scrape inputs. Every morning at 07:00: run inbox loop. Every Sunday evening: run schedule sweep. Three times a day: run Telegram aggregation. End of day: run Notion sync. All write to cos/output/.
6. Daily review surface. Either open cos/output/ directly in Cursor, or read it from your personal dashboard (see #208 case above) where the four daily output files appear as a single panel.
7. Tune weekly. Every Sunday, read the week's output files, ask the agent "where did you mis-classify? what rules should I add?" — and append to AGENTS.md. The system improves itself.

The AGENTS.md (chief-of-staff constitution):

You are my Personal Chief of Staff.

WHO I AM
- A founder / commercial officer at [Company].
- Calendar densest 10:00-19:00 local. Mornings 07:00-10:00 are sacred for deep work.
- Currently focused: [3 priorities, updated monthly]
- Currently NOT engaging: [3-5 topics I'm deferring]

INPUTS YOU CAN READ
- cos/inputs/gmail.json — last 200 threads, oldest-first
- cos/inputs/telegram.json — last 24h of all chats and saved messages
- cos/inputs/calendar.json — next 14 days
- cos/inputs/notion.json — pages I touched since yesterday

OUTPUT RULES
- Always write to cos/output/<YYYY-MM-DD>-<loop>.md
- Lowercase headers. En-dash separator. No emoji.
- Cite source: "[gmail #thread_id]" or "[tg @chat 14:32]" on every claim.
- NEVER invent tasks. If a task isn't explicit in source, write [needs founder confirmation].
- NEVER classify a person as low-priority without 3 prior conflicting signals.
- NEVER touch the calendar without my approval.

DO NOT INTERRUPT
- Anything from [list of 5-10 names — partners, key clients, family]
- Anything tagged [URGENT] or [LP] in subject

INTERRUPT-WORTHY
- Any reply that mentions money figures > $X
- Any message from [list of investors / board]
- Any thread that's been open > 7 days from someone in [list]

WHEN UNCERTAIN
- Ask one specific question. Do not guess.

The inbox loop prompt (cos/loops/inbox/loop.md):

Read cos/inputs/gmail.json + cos/inputs/telegram.json.

For each thread, classify into one of:
- REPLY NOW — needs my hand today, < 5 min response
- REPLY LATER — needs my hand this week, draft a starter response in cos/output/<today>-drafts/<thread_id>.md
- ACKNOWLEDGE — one-tap "got it, will look" reply; you can draft it
- IGNORE — no action needed, archive

For REPLY NOW: output a single-table by-thread with (sender, one-line ask, suggested action, deadline).
For REPLY LATER: write the draft. I will polish and send.
For ACKNOWLEDGE: list them; I'll batch-approve.
For IGNORE: list them silently at the bottom.

End with: "INTERRUPTIONS" section — anything that overrides my DO NOT INTERRUPT list per my AGENTS.md.

Cite every claim. Never invent.

Daily cron (sample crontab):

• Do not let the agent send. Drafts only, ever. The day it sends without your eyes on it is the day a partner gets a wrong number and your trust evaporates.
• DO NOT INVENT TASKS rule (founder CRM lesson from 2026-04 era): explicitly encoded in AGENTS.md. Past pain: agent extrapolated "obvious next steps" and produced a follow-up list that included tasks no one had actually committed to. The founder DM'd the entire fake list to a contact. Hard rule now.
• Tune the DO NOT INTERRUPT list weekly. Stale lists are the failure mode — someone got bumped to "regular" and the agent quietly stopped flagging them.
• Don't run multiple loops concurrently on the same inputs file. Snapshots first, loops second. Otherwise inbox loop and telegram loop write contradictory action items.
• Cost watch: if your loops occasionally blow up to long-context, your Pro plan can be exhausted on a single bad day. Cap per-loop turns and timeout aggressively (--max-turns 20, --timeout 180).
• Time and attention are the only finite resources. Design every CoS surface against the test "does this save me a minute of attention or cost me a minute?" If a CoS feature requires more reading than it replaces, kill it. Operators who skip this test end up curating their CoS instead of being served by it.

• Lighter version (no MCPs yet): start with just an Apple Mail export + a Telegram desktop export dropped in a folder, run the inbox loop manually. You'll discover the pattern before wiring real-time data.
• Heavier version (mid-size agency class): instead of four loops, build a graph that ingests every token continuously (token metabolism). Loops become continuous services. Add a "shadow employee" twin reachable at agent@you.com that handles routine email replies in your voice with delegation rules.
• Team variant — Champions: each team lead runs their own CoS with their own rules; share only the patterns (which loops you found valuable), never the contents. Federation, not template.
• Cron-less variant: instead of cron, invoke loops from your dashboard — "refresh CoS" button calls the same prompts on demand. Lower latency, more energy on you.