Gdpr

gdpr

Overview

Practical GDPR/privacy for my B2C PWAs — not legal advice, an engineering checklist. First classify the app's data posture, then run the decision gates, then apply the guidance for the gaps. Security controls (RLS, secrets, open-write) live in security-bar — referenced, not restated. This skill owns privacy/compliance: lawful basis, minimization, retention, data-subject rights, processors, transfers, the privacy notice, and analytics.

When to use

• Building or shipping a feature that collects/stores/transmits personal data.
• Auditing an existing app for GDPR gaps, or writing its privacy notice.
• Choosing/adding analytics (read the analytics section first — it's coupled to consent).

Step 1 — classify the data posture

My fleet splits three ways; the posture decides how much applies:

1. Offline / on-device (e.g. kartaak, converthub) — no server, no egress of user data. Privacy by architecture. GDPR surface ≈ minimal; the win is stating it (a "runs on your device, we collect nothing" notice = trust + the easy compliance win).
2. Server-backed (e.g. EventSplit, Monete — Supabase) — I'm a controller. Personal data of users and third parties (e.g. friends added to an event who never interacted with me). This is the real surface: lawful basis, minimization, retention, DSR, RLS.
3. Client-side but ships to third-party processors (e.g. mintza → OpenAI/Anthropic/ Google/Azure) — heaviest: multiple processors, international transfer (US), and possibly special-category content. Controller/processor split depends on BYO-key vs hosted — decide it explicitly.

Step 2 — decision gates

Run these on the data path (adapted from ECC's HIPAA gates):

1. Is this personal data? (identifies a person directly or combined — name, email, expense tied to a named person, voice). If no → stop, nothing here applies.
2. Lawful basis? Name it: consent, contract, or legitimate interest (third-party names in a shared event = legitimate interest; document it).
3. Does this processor / model provider need a DPA before touching the data? (Supabase, any analytics tool, any AI API.) International transfer covered (SCCs / adequacy / DPF)?
4. Minimum necessary? Store the least that works (a display label, not full identity).
5. DSR / erasure handled? Can a person get their data deleted? For no-login apps, "delete the event" is the erasure path — make it real and reachable.
6. Auditable? Only where proportionate — see the non-goal below.

Step 3 — guidance for the gaps

• Privacy notice — every app gets one (even offline: "we collect nothing"). Use references/privacy-notice.md. Disclose every processor (Supabase, analytics, AI APIs) and any transfer.
• Retention — cap growing personal data; EventSplit's history/trash caps already act as de-facto retention — document them as the policy.
• Tag PII at the schema level (Supabase posture): COMMENT ON COLUMN events.participant_name IS 'PII: name'; — makes personal-data columns explicit for audits and future migrations.
• Security controls (RLS, open-write, secrets, egress) → eskills:security-bar. Note: EventSplit's open-write RLS means anyone can read/delete personal data → that's a privacy exposure, fix per security-bar.

Analytics — privacy-first, no consent banner

No tracking ⇒ no cookie banner. Don't add one you don't need. To get metrics and keep that win:

• Use cookieless, no-PII, EU-hosted analytics — Plausible / Umami / PostHog-EU (cookieless). No cookie/localStorage identifier, no cross-site → no consent required.
• Do NOT use GA4 / Meta Pixel / anything cookie- or fingerprint-based → that requires consent + a banner (the bloat we avoid).
• Per posture: offline apps can still ship cookieless pageviews; Supabase apps must disclose the analytics tool as a processor; mintza already ships to AI, so cookieless analytics adds negligible marginal exposure.

Never expose PII in… (guardrail list)

Never put personal data in: logs, analytics events, crash/error reports, LLM prompts, URLs/query strings, localStorage/sessionStorage, or screenshots. Prefer opaque IDs over names/email. Never ship the service_role key client-side (→ security-bar).

// BAD — leaks PII into the error (and the error tracker)
throw new Error(
  `No member ${member.name} <${member.email}> in event ${event.title}`,
);
// GOOD — generic message; details logged server-side with opaque IDs only
logger.error("member lookup failed", { memberId: member.id, eventId });
throw new Error("Member not found");

For the AI-processor posture (mintza): send the model the minimum content needed, never attach identifiers you don't have to, and disclose the provider + transfer in the notice.

Not in scope here

• Security controls (RLS, secrets, authz, egress) → security-bar.
• Concrete failure recipes (open-RLS exposure, undisclosed transfer) → stack-gotchas.
• Non-goal (arch-bar): no HIPAA-grade audit-everything, session-timeout, or facility-RLS machinery — disproportionate for no-login B2C PWAs and fights minimization + the egress limit. Audit only where proportionate.
• Scope is EU GDPR; UK-GDPR is near-identical; CCPA is out of scope.