Setup & Customization Checklist

1. Trigger setup

This skill runs from a scheduled trigger (not a webhook), one per team member whose DMs are scanned. Reference trigger instructions:

You are running as [Trigger Owner]. Scan [Owner]'s LinkedIn DMs for
business-relevant signals.

Load and follow the <LinkedIn DM Signal Classifier> skill in full.

Context for this run:
- Trigger owner: [Name] ([email])
- LinkedIn account: [Owner]'s connected LinkedIn
- Run this daily
- Chat limit for this run: 50

• Trigger owner's LinkedIn account is connected to the workspace.
• Schedule: daily is the tested cadence. More frequent runs mostly hit the cursor and do nothing; less frequent risks stale ⏰ awaiting-reply flags.
• One trigger + one skill instance per person — cursors and dedup lists live in the trigger owner's user memory and must not be shared.

2. Placeholders

• {{TRIGGER_OWNER}}, {{CRM}}, {{SIGNALS_CHANNEL}}, {{EXPANSION_CHANNEL}}, {{MQL_CHANNEL}}, {{PARTNERSHIPS_OWNER}}, {{FUNDRAISING_OWNER}}, {{LEAD_SCORING_SKILL}}, {{OUTREACH_SKILL}}, {{HIGH_ACV_THRESHOLD}}, {{SELF_SERVE_THRESHOLD}} — all filled.
• Map the bucket definitions in Step 5 to your funnel stage names (closed-won, open-deal stages, closed-lost/nurture/self-serve). The buckets reference stage semantics, not stage names.
• Optional: seed the trigger owner's personal-connections skip list (Step 3) so friends and peers never generate signal posts.

3. Policy decisions

• Start in EVAL mode. Run with CRM writes disabled for 1–2 weeks and review the "would-have-written" blocks in Slack, then follow the steps as written to go live. This is how the skill was derisked in production.
• Never auto-send — the strongest invariant. The skill may draft a suggested reply for the thread owner and may build approval-queued sequences for other buying-committee members (Step 6.5), but a human sends everything. If you want pure detection with no drafts at all, delete Step 6.5 and the closing-line rule.
• The self-serve gate on PROSPECT_HIGH_ACV — set {{SELF_SERVE_THRESHOLD}} to match where your sales-touch economics actually start; the gate exists because inbound DMs from tiny teams were over-promoting accounts into sales-tier alerts.
• The tier promotion rule (inbound DM promotes one tier up, never down) — confirm it matches your scoring model.
• MQL channel discipline: scored signals post to {{MQL_CHANNEL}} in your standard lead-scoring alert format, never a custom DM layout — confirm the format reference in Step 6.5 points at your scoring skill's alert spec.
• The mandatory end-of-run digest even on zero signals — this is your health check that the scan ran. Don't remove it.

4. Behavioral invariants (do not remove)

• Cursor written only at the end of a successful run — duplicate processing on retry is acceptable, silently skipped signals are not.
• LinkedIn tool calls strictly sequential.
• 7-day per-contact dedup before posting; updates use the short 🔄 DM Update format.
• ⏰ awaiting-reply flag only when the other person spoke last.
• Verbatim last inbound message quoted in every signal post — paraphrase is never a substitute.
• The contact already in the thread is never enrolled in a sequence — reply drafts only.
• Memory-full → skip the write, flag it, never trim existing memory.