Written 2026-11-02, from reviewing a planning-only prompt for making NextVibe white-label.
Nothing is built yet. This is the thinking.
See also: 00 — System Design,
interview: 06, Q11 + S4
The goal
Other businesses run NextVibe under their own brand (domain, logo, colours, emails) on one
codebase and one deployment. NextVibe itself becomes tenant #1.
The decisions, in the order they matter
-
Isolation model
Model Cost Isolation Shared schema + tenantIdcolumnCheapest, one migration path Only as good as your filtering Schema per tenant Medium, migrations run N times Stronger Database per tenant Most expensive Strongest, easy per-tenant backup/restore Shared-schema is the usual start, and the rest of this note assumes it.
-
Enforce isolation centrally, not by memory.
- App layer: a Prisma client extension adds
where tenantId = currentto every query and
stamps it on every write. Weakness:$queryRaw/$executeRawskip extensions completely. - Database layer: Postgres Row-Level Security. Policies check
current_setting('app.tenant_id'). Weakness: the setting has to be applied per transaction
(set_config(…, true)), and connection pooling makes “per request” harder. - Best answer: both. RLS is the backstop for the queries the extension misses.
- App layer: a Prisma client extension adds
-
Tenant context. Carry it with AsyncLocalStorage (
nestjs-cls) rather than request-scoped
providers, so it reaches services, queue jobs, cron jobs and websocket handlers without the
performance cost. -
Resolving the tenant: custom domain (Host header), then subdomain, then an explicit header
for mobile.- The header must be checked against the tenant in the JWT, or a user of tenant A could
send tenant B’s id. - Inbound callbacks have no tenant Host: payment webhooks, OAuth callbacks, email links.
With per-tenant payment keys, you need the tenant before you can check the webhook
signature, so it must come from the payload or a per-tenant webhook URL.
- The header must be checked against the tenant in the JWT, or a user of tenant A could
-
Identity is a product decision: per-tenant users (simpler, safer) or one global identity
with tenant memberships (one person, many branded apps). -
Everything outside the database needs the tenant too: cache keys, queue jobs, file storage
paths, websocket rooms, logs. -
Migration: create tenant #1, then add a nullable
tenantId, then backfill, then NOT NULL +
indexes, then make unique constraints per-tenant (@@unique([tenantId, email])). Each step
needs a way to roll back.
Requirements that are easy to forget
- Offboarding: export and delete one tenant’s data.
- Restoring one tenant from backup without touching the others.
- Per-tenant rate limits, so one noisy tenant can’t slow down everyone.
- Data-protection law in the tenant’s country (for Nigerian clients, NDPR/NDPA).
- Existing money flows (affiliates, organizer billing, payouts) colliding with per-tenant payment
credentials and a platform fee.
What the prompt review taught about writing prompts for agents
- Too big for one pass: split it into discovery, then plan, with your answers in between.
“Ask me before assuming” can’t work inside “give me one report”, so tell the agent to mark
assumptions and keep going. - Don’t prescribe the solution before discovery. “Don’t assume the stack” plus “use an ORM
subscriber” contradicts itself, and steers the agent past the RLS question. - Put YAGNI first: custom-domain SSL, mobile build pipelines and revenue share can wait for the
first real client. - Fix existing security debt as phase 0, before building isolation on top of it.
Open questions (product owner’s to answer)
- Is there a first client lined up, or is this preparation?
- Can one person have accounts on several tenants’ apps?
Software engineering
Infrastructure & security
Computer science
Hardware & engineering