Draft an outreach reply

When a prospect replies to an outreach email, that reply lands as an inbound message on the conversation we created from the approved initial. The AI runner is configured to skip these conversations (agentMode='draft_only'), so without this curator the conversation just sits there. Your job is to read the thread, draft a sensible reply, and file it as a kind: 'reply' proposal for human approval — same Review tab, same approve/dismiss flow as initials.

This is fundamentally different from skill://outreach/draft-initial-email:

• Initial is generative: campaign brief + KB context → personalized first email per contact.
• Reply is reactive: read the inbound message + thread context → addressed-to-the-message reply.

Replies must answer what the prospect actually said. If they asked a question, answer it. If they declined, acknowledge graciously and (if the campaign brief has a soft alternative ask) offer it once. If they want to talk to a human, your reply should suggest a handoff path; the operator approving will route accordingly.

TL;DR

1. Read the conversation with conv_get_conversation(<conversationId>) (the user prompt names it). Note the outreachCampaignId.
2. Read the campaign with outreach_get_campaign(<campaignId>) for the brief — the reply should stay on-message.
3. Identify the latest end-user message — that's what you're replying TO. Earlier messages are context only.
4. Decide intent: question, decline, ask-for-human, off-topic, low-quality, etc. The right reply differs sharply by intent.
5. Ground product claims with kb_search if the prospect asked something factual.
6. Draft a 30–120-word reply that addresses the inbound. Plain prose, no headings, no JSON-escaping. No unsubscribe footer — replies thread inside the existing email chain which already carries the original link.
7. File with outreach_propose_reply({ conversationId, draftBody, evidence }). The proposal lands pending; the operator approves on /dashboard/inbox. Approving sends via conv_send_message.

Step 1 — read the conversation

{ "name": "conv_get_conversation", "arguments": { "id": "ccv_…" } }

The response includes outreachCampaignId, messages[], agentMode. Verify agentMode === 'draft_only' (otherwise the trigger fired in error and you should stop). Verify outreachCampaignId !== null (same — should not have been triggered).

Step 2 — read the campaign

{ "name": "outreach_get_campaign", "arguments": { "id": "ocmp_…" } }

The brief tells you what the campaign is about. Stay on-message. If the inbound veered off-topic ("you wouldn't believe my weekend"), the reply should politely redirect to the original ask, not engage the off-topic detail.

Step 3 — what to skip

• A pending reply already exists for this conversation — outreach_propose_reply will reject. Run outreach_list_proposals({ status: 'pending', kind: 'reply', campaignId, contactId }) first to dedupe.
• The latest message is internal (internal: true) — that's a staff side-comment, not a prospect reply. Skip.
• The latest message is from authorType: 'user' or 'agent' — that's an outbound message we already sent (or that someone just approved). The trigger should not fire on those, but defend.
• The conversation is closed or staffed (status !== 'open' or assigneeUserId set). The operator is handling it; don't draft on top of them.

Step 4 — read intent

• Question: "What pricing tier covers <thing>?" → answer concretely from KB; if KB doesn't cover it, acknowledge and propose a follow-up (call, email, or handoff to a human).
• Decline: "Not interested." → one short, gracious sentence. Don't push back. Optionally one open-ended ask only if the brief has a soft alternative.
• Ask for human: "Can we talk to someone?" → "Of course — I'll have someone reach out today" + thank them. The operator approving the reply is implicitly that handoff.
• Off-topic / pleasantry: brief friendly response then redirect to the original ask.
• Hostile / unsubscribe-intent: don't engage. Draft a one-line acknowledgement. Operator should likely dismiss the draft AND mark the contact suppressed.

Step 5 — ground product claims

If the prospect asked something specific that needs a factual answer ("does X integrate with Y?"), do a kb_search and ground in what comes back. Don't invent. If the KB has nothing, write at the level of "we'd love to walk you through X — when's good for a 15-minute call?" — punt to the operator with a defensible position.

Step 6 — draft

• 30–120 words. Replies are short. Conciseness signals respect for their time.
• Plain prose. No headings, no markdown structure, no bullet lists unless 2–3 items genuinely cluster (e.g. listing product capabilities they asked about).
• Address what they said. First sentence acknowledges or answers; second offers a next step if appropriate.
• No unsubscribe footer. The original initial carries the (signed, surviving-forwarding) link. Replies inside the same thread don't need a duplicate; including one looks robotic.
• Voice: warm, second-person, the way an operator would write if they had time.

Step 7 — file the proposal

{
  "name": "outreach_propose_reply",
  "arguments": {
    "conversationId": "ccv_…",
    "draftBody": "Hi Jane,\n\nGreat question — yes, we integrate with Slack via OAuth (one-click; takes ~30 seconds). I'd be happy to walk you through it on a 15-minute call this week. Would Tuesday or Thursday work better?\n\n— Munin",
    "evidence": {
      "intent": "question_about_integration",
      "kbDocIds": ["kdoc_slack_integration_overview"],
      "reasoning": "Prospect asked about Slack support; KB confirms; offering a quick demo is on-brief."
    }
  }
}

Behavior:

• The proposal lands in pending status, visible to the operator on /dashboard/inbox Outreach tab with a Reply badge.
• An outreach.proposal.created realtime event fires.
• Approving sends the body verbatim via conv_send_message on the same conversation. No unsubscribe footer is added.

What NOT to do

• Don't auto-send. No conv_send_message from this skill. Drafts only.
• Don't ignore the inbound. A reply that doesn't address what the prospect actually said reads as either a bot or a careless human.
• Don't push when they declined. "Just one more thing" is the cardinal sin of cold outreach. One acknowledgement, one optional soft ask, done.
• Don't include an unsubscribe footer. Initials carry it; replies thread inside.
• Don't draft on a closed / staffed conversation. A human is handling it.
• Don't propose multiple replies per conversation. Idempotency on (campaign, contact, kind=reply, status=pending) will reject.

Related

• skill://outreach/draft-initial-email — the symmetric pattern for first-touch emails.
• skill://kb/review-content — the original "drafted candidates, human approves" pattern this skill follows.
• agentMode on conv_conversations — 'draft_only' is what gates the AI runner from auto-replying. Outreach conversations get this set automatically when their initial is approved.