add-person

Add or update a person in noticed. The primary decision is create vs. update: always check whether someone already exists before minting a new record — updating in place is better than a duplicate every time.

Slow and deliberate on identity over fast and guessing.

flow

For each person in the message:

1. Pull out identifiers + context. Names, URLs (LinkedIn / GitHub / Twitter), handles, freeform words. Many people in one message → one batch.

2. Normalize companies first. Web search to canonicalize ("Amplo Market" → Amplemarket). Carry the canonical name into all searches.

3. Resolve identity — this is step one, not an optional nicety.

   • Parallel own-network searches: name-only, and name + canonical company.
   • Multi-word names join with AND in q ("Austin AND Rief", not "Austin Rief") — q OR-matches all fields and explodes on common terms. Single-token names stay as-is.
   • Structured-filter zero = thin data, not absence. If company returns 0, drop it and rely on the name-only result.
   • If still 0 in own network AND the input was a bare name (no URL/handle), web search for a LinkedIn/GitHub + role/company. One search for a straightforward enrich; two reformulations max — then ask. Don't grind.

   Always: own network → web → ask. No public-scope search.

4. Decide per person:

   • Strong match (single hit, name + company aligns, or URL/handle exact) → already in network; update record, log interaction, save memory.
   • Uncertain (multiple same-name hits, no company corroboration) → surface candidates in the preview; don't decide unilaterally.
   • Nothing matches → new contact. With a URL/handle, proceed to create. Bare name → ask first (voice notes mishear names).

5. Determine interaction date. The skill logs a met interaction by default. The interaction date is often not today — it's when the meeting happened, not when the capture was sent. Determine it before writing:

   • Event link present (Luma / Partiful / lu.ma URL, named event with a known date) → use that event's date as occurred_at.
   • Explicit date signal in the message ("yesterday", "at dinner last Tuesday", "at [event] on June 21") → derive the ISO date.
   • No date signal → ask in the preview. Do not silently default to today when the date is genuinely unknown.

   Surface the inferred date in the preview ("logging a met on 2026-06-22 from the Luma link — correct?") so the user can correct it in the same reply.

6. Tag silently. Role, company, sector, event-as-context (e.g. myosin-dinner), location. Applied at write time. User corrects at preview. Tags merge; removal needs the UI.

7. Preview (structure below).

8. On the go-ahead, fan out writes, then read back. Identity questions and date questions both resolve into the same go-ahead reply — not separate round-trips.

---

log interaction — default ON

Log a met interaction for every person added or updated through this skill unless context clearly says otherwise — e.g. the user is pre-loading someone they haven't met yet, or explicitly says "haven't met them." When in doubt, log it.

Dedupe guard: if the person already exists, call get_person(include: 'dossier') and check their interactions. If a met for the same date is already logged, skip — don't double-log.

noticed.log_interaction({
  person_id,
  kind: "met",
  occurred_at: "<ISO timestamp>",      // e.g. "2026-06-22T00:00:00Z"
                                        // omit only if user confirmed "haven't met them"
  payload: { summary: "<one-line what you met about>" }
})

occurred_at is an ISO timestamp. Omitting it silently defaults to now — which is wrong for back-dated events. Always set it explicitly.

---

save memory — rich context here, not in default_notes

After the interaction, save the freeform context — who they are, what they need, what was discussed, how they can help — as a durable memory attached to the person record.

noticed.add_memory({
  person_id,
  content: "[mcp · skill:add-person · YYYY-MM-DD]\n<source-tagged context>",
  tags: [...]      // optional; inherit from person tags where relevant
})

• Each content line tagged [from user] or [research, unverified]. Flag conflicts between the user's account and web research — write both, tag each.
• The date in the memory prefix (YYYY-MM-DD) should match occurred_at of the interaction. They describe the same moment.
• add_memory attaches to relationship.memories on the person record, readable back via get_person. This is the right home for dated, episodic context.
• default_notes is not the right place for rich context. Reserve it for terse, stable record-level annotations (e.g. a pronunciation note, a permanent caveat) — not for "what we talked about at that dinner."
• Tighter memory ↔ interaction linking (foreign-key refs) is a future enhancement once the MCP supports it. For now: keep dates consistent and don't try to force a link manually.

---

preview structure

Organized by what's actionable, not by person. Empty zones collapse (a clean batch is just the ready list + the close).

ready to save to noticed:

<a sentence or short paragraph per resolved person — who they are, what's about to happen,
and what date the `met` will be logged on. One line when context is thin; a few sentences
when there's something to say. Don't pad; never invent detail. Always surface the
interaction date: "logging a `met` on 2026-06-22 from the Luma link".>

need from you:

<only people needing a decision — ambiguous identity, bare name with no corroboration,
unresolved interaction date. Pull each out so it can't be missed. One question per person.>

---
<closing line naming what the reply triggers — "reply with the date and I'll save all three
to noticed" — or just "save to noticed?" when nothing's blocking.>

• Prose, not plan rows. No [merged] / [new] / [from network] tags shown to the user.
• Plain status language: "new contact", "already in your network", "couldn't place her". Never bracket labels or "auto-resolve".
• Names in bold. Lowercase, warm.
• Don't print a tags: row. Fold a tag into prose only if worth surfacing ("saving her as a founder in consumer AI"); otherwise apply silently.
• The close names the effect and says "noticed".

---

provenance — two surfaces

• Chat (preview + readback): web-found facts attributed softly in prose ("his LinkedIn says founder in consumer AI") so research never reads as user-stated. No bracket tags shown in chat.
• Stored memory (system-only): every line tagged [from user] or [research, unverified], written into add_memory content. Never shown verbatim in chat.

Flag conflicts between the user's account and research — write both, tag each.

---

readback (after writes land)

Recap what happened per person — unprompted. noticed writes are silent; this is the only confirmation. Write like a person, not a database row: who's new vs updated, what interaction was logged and on what date, the memory gist, the link if any. Flag web-sourced facts conversationally ("her headline's from the web, worth a glance").

---

write sequence

Structured fields (linkedin_username, headline, github) can only be set at creation via add_to_network; update_person cannot write them. Web-found headline → free_form.headline on a new record, or the memory body ([research, unverified]) on an existing one.

add_to_network's default_notes param silently drops — always set notes via a follow-up add_memory call, not on the creation call.

# — already in network —
noticed.get_person({ person_id, include: "dossier" })
  # check for existing met on this date before writing
noticed.log_interaction({ person_id, kind: "met", occurred_at: "...", payload: { summary: "..." } })
noticed.add_memory({ person_id, content: "[mcp · skill:add-person · YYYY-MM-DD]\n<source-tagged context>" })
noticed.update_person({ person_id, tags: [...] })          # tags only; not default_notes for rich context

# — new contact —
noticed.add_to_network({ free_form: { name, linkedin_url?, github_login?, headline? }, tags: [...] })
  # → person_id returned
noticed.log_interaction({ person_id, kind: "met", occurred_at: "...", payload: { summary: "..." } })
noticed.add_memory({ person_id, content: "[mcp · skill:add-person · YYYY-MM-DD]\n<source-tagged context>" })

---

invariants

• Never fabricate structured identity fields. Blank beats a plausible guess. LinkedIn URL, company, handle only from explicit input or confirmed web source.
• Search own network before add_to_network. One duplicate is worse than one extra lookup.
• Preview before save. One confirm per batch. Corrections fold into the same reply. Always read back.
• log_interaction is append-only. Check existing interactions via get_person(include: 'dossier') before writing; don't double-log the same event on the same date.
• No writes before the go-ahead.

---

tool call reference

noticed.search_people({ q: "Name AND Company", scope: "own" })
noticed.get_person({ person_id, include: "dossier" })
noticed.add_to_network({ free_form: { name, linkedin_url?, github_login?, headline? }, tags: [...] })
noticed.update_person({ person_id, tags: [...] })
noticed.log_interaction({ person_id, kind: "met", occurred_at?: "ISO timestamp", payload?: { summary } })
noticed.add_memory({ person_id, content: "...", tags?: [...] })
web_search(...)     // canonicalize companies, enrich thin identity

explicitly NOT in scope

• scope: "public" searches (that's search-network's job)
• Writing default_notes for dated or episodic context (use add_memory)
• Surfacing provenance tags or a raw tags: row in chat
• Logging an interaction when context clearly says the user hasn't met this person yet
• Writing before the go-ahead