| name | data-minimization-designer |
| description | Design systems to collect, transfer, process, and retain less data — payload slimming, field-level collection review, sampling strategies, aggregation-at-source, and retention-by-design. Use this skill whenever the user designs telemetry/analytics/logging pipelines, asks what data to collect, mentions data hoarding or GDPR-style minimization, or builds APIs whose payloads carry unused fields. Part of Lean Agentic AI Skills; emits lean-findings.json plus a minimization design. |
Data Minimization Designer
Advisor skill. Input: data flows (what's collected, transferred, stored, and actually used), pipeline configs, API schemas. Output: minimization design + lean-findings.json.
Every unnecessary byte is charged four times: collected (client E), transferred (network E), processed (pipeline E), stored (storage E+M) — and it compounds with retention. Minimization is the only pattern that reduces all four at once, and it doubles as privacy hygiene.
Subject type: emit subject.type: "architecture" in findings.
Design moves
- Collect for questions, not "just in case". Field-level review: for each collected field, which query/dashboard/model consumes it? Unconsumed fields (verify against actual query logs where possible) are high-severity findings. New pipelines: start minimal, add fields when a question demands them — adding is easy, deleting is politics.
- Sample where aggregate truth suffices — performance telemetry and high-volume traces at 1–10% sampling with correct weighting; keep 100% only for errors and audit-mandated events. Name the trade-off: rare-event visibility drops with sampling; keep errors unsampled.
- Aggregate at source — send histograms/counters, not raw events, when only aggregates are consumed (metrics vs events distinction). Massive transfer/storage reduction; loses drill-down — say which questions become unanswerable.
- Slim payloads — API responses with fields no client reads (pair with db-efficiency-audit's SELECT-* findings), sparse/columnar formats for analytics (Parquet over JSON), compression at every hop.
- Retention by design — every dataset gets a TTL at creation, tied to the question it serves (debug logs: days–weeks; aggregates: long). Pairs with storage-lifecycle-audit for the brownfield version.
- Derive, don't duplicate — one raw source of truth, materialized views for consumers, not N copies.
Honesty rules
Minimization can destroy future optionality — the fix for that is a documented decision ("we chose not to collect X because no current question needs it"), not silent hoarding. Legal/audit retention mandates override TTL preferences; always carve them out. Compliance framing (GDPR data minimization) is a bonus argument, not legal advice — say so.
Not this skill's job
Tiering data already stored (storage-lifecycle-audit), query efficiency (db-efficiency-audit).
