Run Your Professional Network as an AI Project

Start with one small area for the standing responsibility called "network".

network/
  AGENTS.md
  inbox/
  workflows/
    discover-opportunities.md
    log-contact.md
    plan-follow-up.md
  contacts/
  events/
  decisions/
  archive/

AGENTS.md routes work. The inbox holds raw material before review. Contacts and events own checked facts. Decisions preserve why a rule exists. The archive holds completed work that still has a justified reason to be retained.

This adapts the structure from How to Structure AI Projects With PARA. The folder shape is useful because it tells the agent where it may look and what it may change.

Give the agent four preparation permissions:

• discover approved opportunities
• summarise checked material
• draft possible follow-ups
• remind you about an explicit date

Reserve four actions for yourself:

• send
• apply
• reply
• accept

An outbound action can affect trust, disclose context, or create a commitment. Anthropic's agent framework describes the same wider principle: people should retain control, particularly before high-stakes decisions are made.

Put notes, links, messages, and event pages into the inbox before filing them. Inbox content is unverified. It should not silently become a contact fact or deadline.

During triage, move a checked fact into its owner file, create one follow-up item, archive the input, or delete it. Do not copy the same next step into several files.

A claim about a person, company, or event needs the date it was checked and the evidence checked. Without both, keep it as an unverified note.

Use a generic shape rather than one large profile:

record_id:
relationship_context:
organisation:
source_checked:
checked_at:
last_interaction_date:
what_was_said:
commitment:
next_review_date:
status:

These are possible field names, not a request to fill every field. Record only what serves the system's stated purpose.

European Commission guidance on GDPR principles covers data minimisation, accuracy, storage limitation, integrity, confidentiality, and accountability. This guide is an operational pattern, not legal advice.

The agent may search approved sources, summarise requirements, and prepare a dated opportunity note.

• What source was checked?
• When was it checked?
• What deadline does the source state?
• Which requirement is still unclear?

The agent may draft an application outline. It may not apply, accept terms, or represent your interest.

The agent may turn approved notes into the generic record shape. It must preserve the source and checked date.

If the input does not support a field, leave it empty or mark it unverified. Do not let the agent enrich a record from unrelated public sources by default. More personal data is not automatically more useful.

The agent may identify an explicit promise, calculate a review date from a written rule, and draft a message. The draft stays in a review queue.

You decide whether the context still supports contact. You edit and send the message yourself. The record may note the action only after it actually happened.

A private location is necessary, but location alone is not security. Use authentication, encryption, backups, and access limits appropriate to the risk.

The European Data Protection Board's security guidance recommends risk-based controls including access management, backups, encryption, and reviews of authorisations.

Set a recurring review. Correct inaccurate information, remove what is no longer necessary, and archive completed opportunities only when there is still a justified reason to retain them.

Do not build a complete contact-management system on day one. Create the instructions file, one workflow file, and one generic record.

Run one real item from inbox to reviewed follow-up. Add structure only when repeated use proves it is needed.