luma-inbound-webhook-enrich

Solid

DRAFT. Design and harden a standing Luma registration/unsubscribe webhook that enriches guests and writes to CRM or customer DB — with backfill and versioning. Not installed by deepline skills.

AI & Automation 62 stars 12 forks Updated yesterday MIT

Install

View on GitHub

Quality Score: 83/100

Stars 20%
60
Recency 20%
100
Frontmatter 20%
70
Documentation 15%
71
Issue Health 10%
50
License 10%
100
Description 5%
100

Skill Content

# Luma inbound webhook enrich (DRAFT) ## Goal Registration events become a **durable system**: webhook → validate → normalize LinkedIn → enrich → upsert CRM/DB → optional Slack notify. Not a one-off CSV import after the event. ## Design checklist 1. **Trigger:** Luma registration / unsubscribe (document exact event names you subscribe to). 2. **Idempotency:** upsert key = email (plus event id). 3. **Normalize:** LinkedIn slug → canonical `https://www.linkedin.com/in/...` URL. 4. **Enrich:** cheap fields first; gate expensive providers. 5. **Write:** CRM or customer DB fields listed in `references/field-map.md`. 6. **Failure mode:** keep original failed payload; dead-letter queue; backfill later. 7. **Versioning:** bump play/workflow version on each mapping change; never silent overwrite of prod without dry-run. ## Backfill pattern 1. Export failed webhook payloads (or Luma guest delta since last success). 2. Re-run through the **current** versioned play. 3. Leave original failed records intact for audit. 4. Report: attempted / succeeded / still failing. ## Pair with - `luma-event-campaign` for outbound staging - `event-ops-score-and-door` for capacity decisions (separate from inbound enrich) ## Guardrails - Do not log raw PII into public git. - Secrets only via secret store / workflow secrets — never in SKILL.md. - Ask before replaying large historical backfills (cost + CRM noise).

Details

Author
getaero-io
Repository
getaero-io/gtm-eng-skills
Created
7 months ago
Last Updated
yesterday
Language
TypeScript
License
MIT

Similar Skills

Semantically similar based on skill content — not just same category