follow-up

follow-up

Draft a follow-up to one person the user recently met IRL. One person at a time. Resolve identity, channel, and angle before writing a word.

Person resolution is shared with remember-person — same own → ask order. No web enrichment, no public scope.

flow

1. Resolve the person. search_people own, multi-word names AND-joined. Strong match → use it. Multiple/no match → surface candidates, never default to new silently. On "new", get a linkedin before drafting so they can be added after the send is confirmed.

2. Read recent context. get_person({ include: "dossier" }). If nothing logged in ~2 weeks and notes are thin, ask "what was the meeting about?" and wait.

3. Pick the channel from what's on the record (linkedin_username, a stored email, a stored phone/handle). One usable → use it. Multiple → ask. Nothing stored → ask how they want to reach them. If linkedin, resolve the mode first (it sets the char budget).

4. Suggest an angle, don't just ask. Read the context and propose the angle(s) that fit ("looks like the natural next step is sending the deck you mentioned — want that, or a quick thank-you?"). Fall back to an open "what's the angle?" only when there's no context. Wait.

5. Draft:

   • In the user's voice — the agent knows how the user writes; match it, don't impose a house style.
   • Short. DMs under ~80 words; email body under ~150; LinkedIn connection note 300-char hard cap.
   • Reference the meeting context naturally.
   • Always include something actionable — ideally something the user offers to do for the recipient (send the deck, make an intro, share a doc, pick a time). A follow-up with no next move is noise. Never generic "would love to stay connected" filler.
   • No em-dashes, no emojis, no AI-tropes.

channels

All channels are copy/paste. The skill produces text the user sends; it doesn't send anything or depend on a connected mail/DM tool.

• email — produce a subject + body (subject auto-generated from the angle); show both to copy into whatever mail client they use.
• linkedin — two modes, resolve before drafting:
  • connection request note (not yet connected): 300-char hard cap, enforced before drafting; still carries a concrete next move, never empty "let's connect" filler. Show the char count.
  • DM (already connected): longer (~80 words), room for a fuller ask.
  • Pick: in-network/logged or "message/DM X" → DM; just-met/"connect with X" → connection note; unclear → ask one line. Provide the profile link; copy/paste, no pre-fill.
• whatsapp — https://wa.me/<phone>?text=<urlencoded> (pre-fills).
• telegram — https://t.me/<username> (no pre-fill).
• sms — sms:<phone>?body=<urlencoded> (mobile).

logging

Never log until the user confirms "sent it". Then:

• Email → log_interaction(kind: "emailed", payload: { channel, angle, summary }).
• DM → log_interaction(kind: "messaged", payload: { channel, angle, summary }); for LinkedIn note the mode in the summary so the timeline shows whether the user is now connected.

Brand-new contact: add_to_network fires with the first log_interaction after "sent it", not before. Then read back what was logged (and added).

No default_notes writes, no memory_save calls — interaction-only.

tool needs

• noticed: search_people, get_person, add_to_network (only on a confirmed brand-new contact), log_interaction
• No Gmail — email is copy/paste like the DMs.

explicitly NOT in scope

• Cold outreach / first-touch; reconnects with dormant contacts; batch follow-ups; sending directly / delayed send (everything is copy/paste); calendar invites; scope: "public" searches; writing to default_notes (interaction-only).