Build Your Personal Assistant Operating System

3.6

Keep People From Becoming Loose Notes

Your assistant already knows who matters to you because it reads your email and calendar

You do not build a contact database by typing in names and birthdays. The assistant observes: who appears in your email, your calendar, your meeting notes. It builds contact records from those interactions. It notices that Sarah Chen has emailed you twelve times in the last month and appears on your calendar every Tuesday. It notices that David Park sent you a birthday message in March and that you have not replied to his last email.

Your job is reviewing and correcting, not entering. When the assistant proposes a contact record from a meeting, you review the fields and approve, correct, or delete before saving. The best data entry is no data entry at all.

For example, if you find yourself typing a client's email address that the assistant already saw in a calendar invite, that is a design failure worth fixing. The same idea applies everywhere: tasks inferred from emails, spending inferred from receipts, energy inferred from journal entries. The relationship is where you first see this extraction at work.

A flat teaching diagram showing email, calendar, and meeting notes reviewed into a contact record and reconnection queue — The assistant builds contact records from interactions you already produce.

Relationship Tracker

Promise	A living contact record for every person who matters, enriched automatically from emails, meetings, and your corrections, with reconnection prompts when relationships go quiet.
Sources	Email interactions, calendar invites, meeting notes, manual entries. Contact records from .
Output	Updated contact records with latest interaction dates, , notes, and a for contacts who have gone quiet beyond their expected cadence.
Saves to memory	Contact records with name, role, , last interaction, interaction cadence, birthday, and notes.
Approval boundary	Reads and updates contact records freely. Reconnection prompts are suggestions only. Never sends messages, schedules meetings, or contacts anyone without your approval.
Failure behavior	If two records appear to describe the same person, flags both for your review rather than merging automatically. If interaction data is ambiguous, marks the record as uncertain.
Review criteria	Contact records should be accurate and current. No two records should describe the same person. Reconnection prompts should reflect genuine relationship priorities rather than surface every contact with missing data.
Graduation trigger	When contact records are stable and reconnection prompts consistently match your priorities, schedule the tracker to update daily and surface the weekly.
Privacy tier	Standard for professional contacts. Personal relationships and any notes containing private circumstances get a higher .

Relationship tiers determine how closely the system watches each connection

You do not track everyone the same way. The system needs tiers, and the tiers determine how aggressively the assistant monitors each relationship.

Relationship tiers and their tracking levels

Tier	Who belongs	What the system tracks	Staleness threshold
Inner circle	Close family, closest friends	Birthdays, gift history, recent conversations, personal	Two weeks without interaction
Active professional	Current clients, close colleagues, direct reports	Role, company, last meeting, open commitments, email frequency	Three weeks without interaction
Wider network	Past colleagues, acquaintances, industry contacts	Name, role, last interaction, from most recent conversation	Six months without interaction
Passive	People you have interacted with once or twice	Name and last interaction date only	No staleness tracking

You set the tiers. The assistant proposes assignments based on interaction frequency. You correct: move this person up to inner circle, move that person down to passive. The corrections train the system. After a month, the tiers stabilize.

A flat teaching diagram showing relationship tiers as attention rings with inner circle, active, wider, and passive labels — Tiers tell the assistant how closely to track a relationship and when silence should become visible.

A simple health score shows which relationships need attention

Each contact gets a health score combining: recency of contact, frequency of interaction, whether there are open commitments (things you said you would do and have not), and whether the relationship is trending up or down.

When you review the relationship list and see a 'cooling' score on a client contact, open your email to verify the last interaction date before deciding whether to reconnect. A 'warming' score on a colleague means recent interactions have been frequent and positive. The score is a directional check, not a verdict.

The morning brief reads these scores. When you are meeting someone today whose relationship is 'cooling,' the dossier says so. When a birthday is coming for someone in your inner circle, the flags section mentions it. The relationship feeds the brief with you would not have assembled on your own.

The prioritizes who you should reach out to next

When you open the , each entry shows a name, the days since last contact, and a suggested action. The assistant ranks entries by combining relationship importance with time since last interaction, adjusted for upcoming calendar events.

capture after meetings turns small details into future preparation

The highest-value habit in this is spending thirty seconds after a meeting telling the assistant what you learned about the people you met. 'Alex mentioned his daughter is applying to colleges. Maria is dealing with a team reorg. David said he is training for a marathon.'

These details go into the contact record and surface in the dossier before your next meeting with those people. Six months from now, when you meet Alex again, the dossier says: 'Last time, he mentioned his daughter was applying to colleges. Ask how that went.' This kind of preparation makes people feel valued, and it costs thirty seconds per meeting.

A flat teaching sequence showing a meeting detail captured into a contact record and resurfaced in a dossier before the next meeting — A detail captured once becomes useful only when it returns at the next moment of preparation.

The gift and gesture layer remembers preferences so you do not have to

For inner-circle contacts, the system tracks mentioned preferences: 'She said she is into pottery right now.' 'He mentioned wanting to try that new restaurant.' When a birthday or holiday approaches, the system surfaces these mentions as suggestions.

Privacy boundaries determine what the system tracks and what it ignores

For example, you might tell the assistant to save meeting notes and email interactions for professional contacts, store only names and birthdays for personal relationships, and never record health-related details beyond a flag that the topic came up. Never auto-import contacts you have only emailed once. These boundaries are set in the before you build.

These boundaries are part of the . The assistant follows them. If you do not set them, the assistant defaults to tracking everything it can infer, which may feel helpful or invasive depending on your comfort level. Set the boundaries before you build.

How this module breaks

What goes wrong	How you notice	What to correct once	What rule to save
The assistant creates two records for the same person because their name appears differently across sources.	Your contact list shows 'Mike Johnson' from email and 'Michael Johnson' from calendar invites as separate people, splitting their interaction history.	Merge the records manually and add a matching rule for this person. Tell the assistant to check for name variants before creating new contacts.	Before creating a new contact record, check existing records for name variants (first name/full name, nicknames, maiden names). Flag potential matches for review rather than creating duplicates.
The assistant infers relationship that feels intrusive or presumptuous.	A contact note says 'seems stressed about the merger' based on email tone analysis, which feels like surveillance rather than helpful .	Restrict relationship notes to factual observations: what was discussed, what was committed, what is upcoming. Remove analysis from contact records.	Contact notes record facts: topics discussed, commitments made, upcoming events. , mood, or emotional state analysis is excluded from contact records.
Stale relationship details persist because no updates old records.	The morning brief introduces someone by their role from two years ago because the contact record was never updated.	Add a staleness rule: contact records not updated in 90 days carry a staleness flag. Modules that reference stale contacts note the age of the information.	Contact records not updated in 90 days are flagged as potentially stale. Modules referencing stale records include a note: 'Last updated [date]; verify current role and .'

Your turn

Build a relationship tracker from your existing interactions

You'll produce enriched contact records for five real people with interaction history, relationship health, and reconnection suggestions.

Why this exercise matters

Relationship intelligence from existing interactions gives you preparation for meetings and prevents important connections from going cold without you noticing.

You’ll leave with

Enriched contact records for five real people from your email and calendar data.
Relationship tiers assigned to each contact.
A relationship health score based on recency, frequency, and open commitments.
A reconnection queue with specific suggested actions.
Privacy boundaries for what the system tracks.

Use the prompt in order

1
Let the assistant scan your email and calendar and propose contact records
The assistant reads your recent emails and calendar events and proposes contact records for the people who appear most. Review the proposals. Correct names, roles, and relationship tiers.
2
Set your privacy boundaries for each tier
Tell the assistant what level of detail to track for each tier. Inner circle gets birthdays and personal . Professional contacts get role and interaction history. Passive contacts get name and last interaction only.
3
Review the relationship health scores
The assistant computes a health score for each contact. Check whether the scores feel right. If someone is marked 'cooling' but you just spoke with them outside of email, correct the interaction record.
4
Act on one reconnection suggestion
Pick one person from the and follow through. Send the message, make the call, or draft the email. One real action validates the better than reviewing ten suggestions.

Prompt details

Each detail will be inferred, or you will be asked to clarify in the chat.

2/2 details

2 details will be inferred, or you will be asked to clarify in the chat

Starter prompt text

Open the full text if you want to check what will be copied.

Show starter prompt

Scan my recent emails and calendar events and build a relationship tracker.

For each person who appears in my recent activity, propose a contact record with: name, role, relationship tier, last interaction date, interaction frequency, open commitments (things I owe them or they owe me), and any personal context mentioned in emails or meetings.

Assign relationship tiers:
Infer from what you know about me or ask if it is unclear

Compute a relationship health score for each contact based on recency, frequency, and open commitments. Flag contacts whose score is 'cooling' or below.

Generate a reconnection queue ranked by importance multiplied by time since last contact, with a specific suggested action for each person.

Privacy boundaries:
Infer from what you know about me or ask if it is unclear

Do not contact anyone on my behalf. Do not share contact details outside this conversation. Suggest actions and wait for my review.

What the answer should give you

Contact records for five people with names, roles, tiers, last interactions, health scores (active, cooling, cold), and a reconnection queue with three suggested actions. Privacy boundaries are applied. Personal context is stored only where the tier allows it.