hey-mantle/mantle-brain/mantle-flows-to-n8n

Mantle Flows → n8n

What this skill does

Helps a partner migrate off Mantle Flows onto n8n (or Make / Zapier). A flow is a graph of triggers → conditions → waits → actions; recreating the boxes-and-arrows is easy. The two parts that actually take work — and that this skill is built around — are:

1. The derived events stream. Mantle computed many triggers on a schedule

("trial expiring in 3 days," "usage crossed the cap," enriched churn reasons). n8n only receives raw webhooks. You must rebuild each derived event as a Schedule-Trigger scan with a fire-once guard, or the dependent flows never fire.

1. In-flight sequences. A customer parked in a "wait N days" step exists only

as a scheduled wake-up inside Mantle and cannot be transferred to n8n. If you cut over without a plan, those customers are silently abandoned mid-sequence.

The full mechanics live in [`references/implementation-spec.md`](./references/implementation-spec.md). Read it before building. This file is the procedure; the spec is the reference. Two ready-to-use artifacts ship alongside it: `references/inventory-template.md` (catalogue your flows first) and `references/n8n-workflow-prompt.md` (generate importable n8n JSON per flow).

When to use it

Trigger on intents like: "migrate my Mantle flows to n8n," "recreate my trial / dunning / win-back automations," "Mantle is shutting down, how do I move my automations," "rebuild the events that trigger my flows," or "turn my flows into n8n workflows."

How to drive the migration

Work through these phases. Inventory before you build — the catalogue is what reveals the real scope (the derived events and the in-flight populations).

Phase 1 — Orient (ask, then read)

1. Confirm the target: n8n (default), Make, or Zapier. The spec's mapping

table covers all three; the generation prompt targets n8n JSON.

1. Find where the partner's flows live today (a Mantle export, screenshots, or a

list). Determine: which events trigger them, what conditions/waits/actions each has.

1. Establish what the partner already has in n8n: an instance, webhook intake,

and an API/DB the flows can query and write back to.

Phase 2 — Read the reference

Read `references/implementation-spec.md` end to end. §0 (the two hard parts), §2 (raw vs derived), and §3 (in-flight cutover) are the non-negotiable core; §7 is the MUST/MUST-NOT checklist.

Phase 3 — Inventory

Catalogue every flow using `references/inventory-template.md`. Per flow record: trigger + trigger kind (raw / derived / scheduled / manual), conditions, ordered steps (waits + actions), the in-flight count (customers parked in a wait right now), and the derived-event dependencies. Then collapse the derived events into a de-duplicated backlog (many flows share one scan).

Phase 4 — Rebuild the events stream (the real work)

For each distinct derived event, build one small n8n workflow: Schedule Trigger → query the partner's data → fire-once guard → continue the flow (spec §2.2–2.3). Build the raw intake too: register Shopify/Stripe webhooks to n8n Webhook nodes, and emit in-app events for app-action triggers.

Phase 5 — Rebuild the flows

One n8n workflow per Mantle flow, via the mapping table (spec §4) and action catalog (spec §5): trigger → conditions (IF/Switch ahead of actions, both branches wired) → waits (Wait nodes) → actions. Write-back actions (tag, set field, extend trial, assign owner, CRM) are HTTP Request nodes to the partner's own API — ask for the endpoint and auth; never invent them. To accelerate, generate each flow with `references/n8n-workflow-prompt.md` and import the JSON.

Phase 6 — Wire and verify

Bind credentials, replace every placeholder (URLs, field names, channels), and test each flow with one real customer end to end (trigger → branch → wait → action). Confirm the fire-once guard prevents re-sends on the next scheduled scan.

Phase 7 — Cut over (the in-flight plan)

For each flow with a wait step, execute the spec §3 plan: count the parked population, pick drain (default) / re-enroll / abandon, stop new Mantle enrollments, run both engines in parallel for one full sequence length, then decommission. Never hard-cut a flow with parked customers.

Guardrails

• Classify every trigger raw vs derived first — the derived set is the

migration's real scope. Re-pointing webhooks is trivial; rebuilding the scans is the project.

• Every derived scan needs a fire-once guard — without it, a daily scan

re-sends to the same customer every day.

• Never strand in-flight customers. Count the parked population per

wait-bearing flow before disabling it. Default to draining; never hard-cut a flow that has people parked in a wait.

• Write-back actions are HTTP calls to the partner's own API — ask for the

endpoint and credentials; do not fabricate them.

• Conditions are branch nodes (IF/Switch) ahead of the actions — and wire

both branches.

• Don't rebuild Mantle's internals (status tables, version snapshots, the

wake-up queue, bulk-run staging). Reproduce behavior, not machinery.

• Generated/imported JSON is a skeleton — credentials, endpoints, and field

names are environment-specific. Verify with a real test before enabling.

• De-duplicate derived scans across flows — build the shared "trial-expiring"

or "usage" scan once.