research-person

research-person

Single person, deep dive. Read everything available, present cleanly, then offer to save the enriched context back to noticed.

Person resolution is shared with remember-person — same own → web → ask order. Read-heavy: the deliverable is a dossier in chat; writes only on an explicit "save".

flow

1. Resolve the person via remember-person's flow: own-scope search, multi-word names AND-joined. Strong match → use it. Multiple hits → surface candidates. Nothing in own → not in network yet; web-only enrichment, mark not-in-noticed. No public-scope search.

2. Read what's in noticed. get_person({ include: "dossier" }) for default_notes, tags, recent log_interactions.

3. Web search to enrich: the person (news, posts, role); their company (canonicalize first; news, funding, product); public output (latest tweet/post/repo). Stop rule: two reformulations that don't converge → ask.

4. Render a scannable dossier in chat. Easy to skim, ideally one screen. Sections (skip empties): overview / interaction history (noticed log_interactions, dated) / company context / conversation starters / key notes / sources. Lowercase, friend-tone, names in bold. (Provenance: see below.)

5. Proactively offer to save. Always close with the offer — don't wait to be asked: "want me to save this back to 's record?" Wait for an explicit "save".

6. On "save": a tight 3-5 line summary of the new findings (not the full dossier), each line source-tagged (system-only). Then read back what was written.

   • In noticed → append a dated [mcp · skill:research-person · YYYY-MM-DD] block; merge tags.
   • Not in noticed → ask "add them to your network too?" Yes → add_to_network + update_person. No → save nothing.

provenance — two surfaces

• Dossier (chat): attribute web findings inline ("linkedin says…", "a recent piece notes…") so research never reads as fact. Surface conflicts with noticed notes rather than smoothing them.
• Stored note (system-only): every saved line tagged [from user] or [research, unverified], written into default_notes, never shown in chat. Since this skill mostly writes research-derived claims, most saved lines are [research, unverified]. (A ~40k-users figure from an aggregator must not sit in the note looking as trusted as a firsthand fact.)

writes (only on "save")

Read existing notes via get_person first; never overwrite. MCP limit: update_person can't set structured identity fields, so a web-found headline for someone already in noticed goes in the note body, tagged [research, unverified].

the rule

• One person at a time. Read-heavy; writes only on "save".
• Never silently default to the wrong identity. Ask when ambiguous.
• Don't grind: two reformulations, then ask.
• On save, read back what landed.

tool needs

• noticed: search_people, get_person, update_person, add_to_network (only when the user wants to add a new person)
• web_search: enrichment + canonicalize companies
• No Gmail (can't assume it's connected).

explicitly NOT in scope

• Batch research; cold profiling without intent; scope: "public" searches; log_interaction (event-debrief's job); auto-save without "save"; surfacing provenance tags in chat; composed plan-handoff (V3).