Workflow 017: Snape, The Relationship Manager

I have an agent whose only job is remembering people.

Every Zoom call. Every email. Anything it can find publicly. Even people I meet in real life.

Snape is my relationship manager. He keeps a record of everyone I've interacted with: who they are, how we met, what we last discussed, and how often we're actually in touch. Before any call or conference, I ask him what my last interaction with that person was, and he tells me.

He's named after the Harry Potter character, like most of my agents. Part spy, part relationship manager.

What this replaces

Relying on your own memory for every relationship you have.

I meet a lot of people. Calls, conferences, introductions, emails. The details blur fast: what someone's working on, what we said we'd follow up about, whether we've spoken twice or twenty times.

The failure isn't forgetting a name. It's showing up to a conversation with nothing, when the other person clearly remembers everything.

Blueprint

Several agents feed one database.

Trigger: automatic, daily, plus manual voice notes for people I meet offline. Inputs: Zoom transcripts, email activity, public information, my own voice notes. Output: a verified record per person, with history, and answers when I ask.

How the pieces move:

• Mrs. Norris collects the day's Zoom transcripts and files them into the CRM
• Snape runs in the morning, checks every entry landed on the right person, and updates records where something significant changed
• Ella, my email agent, sends Snape a report so email activity updates the relationship record too
• For people I meet in person, I send a voice note and it gets added

Organize

What has to exist first:

• One CRM with a record per person. Mine lives in Notion
• Zoom connected, with transcripts saved
• Email connected
• A way to capture people you meet offline, even if it's just talking into your phone
• Identifying details the agent can verify against, usually email addresses

The database is the whole thing. The agents are just what reads and writes to it.

Systematize

Two prompts, because collecting and verifying are different jobs.

THE COLLECTOR

You are my transcript collector. Every day, gather my meeting transcripts
and file them into my CRM.

For each meeting:
1. Pull the transcript
2. Identify who was in the meeting
3. Match each person to their existing record in [CRM], or create one if
   they're new
4. File the transcript against that record with the date
5. Add a short summary of what was discussed

If you can't confidently identify who someone is, file it and flag it
rather than guessing. A wrong match is worse than an unmatched entry.

THE VERIFIER

You are [AGENT NAME], my relationship manager. Every morning, check the
entries that were filed yesterday, then update my CRM.

VERIFY
For each new entry:
- Confirm it was matched to the right person, checking the email address
  they joined with and any other identifying details
- Fix anything filed against the wrong record
- Flag anything you can't resolve

UPDATE
For each person with new activity:
- Update what we last discussed and when
- Add anything significant I learned about them or their business
- Note how the relationship stands now

You'll also receive reports from my email agent. Fold that activity into
the same records.

ANSWER
When I ask about someone, tell me:
- Who they are and how we know each other
- The last time we interacted and what about
- Anything relevant about their life or business right now
- How often we're actually in touch

Never guess. If a record is thin, say so.

How to set it up

1. Build the CRM first, one record per person.
2. Connect Zoom and email.
3. Build the collector and let it run for a week, watching what it files.
4. Build the verifier separately, and check its corrections before trusting them.
5. Get in the habit of voice-noting people you meet offline. That's the manual part and it stays manual.

What about people I only meet in person?

There's less to work with. No transcript, no email history, no ongoing record. So when I meet someone I want to remember, I voice-note it: who they are, what we talked about, why they matter.

It's thinner than a Zoom record. It's also far better than nothing, and it fills out properly the first time we have a call.

Assign

What the agents do: collect, match, verify, update, and answer questions.

What stays human: everything that reaches the person.

Snape tells me what someone's working on and when we last spoke. He doesn't message them. He doesn't send anything. When it's someone's birthday and I ask what gift would actually mean something to them, he gives me context from what I've learned about them. Choosing it is mine.

That line is the whole point. I like giving gifts that matter, and a gift chosen by an agent is not a gift that matters. What the agent removes is the part where I fail to remember the birthday at all.

Rule of thumb from this build: let AI hold the history, keep the human in the relationship.

There's a verification detail worth stealing too. Snape's first job every morning is checking that yesterday's entries went to the right person. In a system that files automatically, wrong data is more expensive than missing data, because you'll act on it confidently.

Integrate

This is the workflow where the stack becomes visible:

• Mrs. Norris collects transcripts and files them
• Snape verifies the entries and updates the records
• Ella reports email activity into the same records
• The meeting agent reads those records to brief me before calls
• Which means every call improves the record that prepares me for the next one

Four agents, one database, no maintenance.

The reason it works is that each agent has one job. Collecting, verifying, and handling email are different problems. Splitting them means I can fix one without breaking the others.

Tomorrow: the research agent that briefs me on people I haven't met yet.