lookout

lookout

!checklist init ${CLAUDE_SKILL_DIR} --force

A lookout watches the real horizon — the live sea, not the chart — and is also the conscience posted to call danger before the ship runs into it. lookout is the sixth skill of the surface suite, the delivery-and-observability lens, and it carries the suite's ethics across the line from philosophy into a launch gate. Its mindset is one shift: shipping is not the finish line, it is the start of the conversation with reality. Its product is a delivery-and-observability setup — preview per PR, RUM and error tracking, an analytics-and-experiment framework — plus the pre-launch ethics gate that the objective function from bearings must pass. It runs across gated stages and will not advance past a GATE until the checklist tool clears it — order enforced, substance yours.

The heavy part, stated up front: the moment the metrics dashboard exists, it starts pulling every decision — so the ethics must already be in place. A/B testing is, mathematically, a gradient-descent optimizer: each experiment keeps the higher-scoring variant, and run thousands of times it slides down the gradient of whatever metric you chose. Point that optimizer at an engagement metric — time-on-site, scroll depth — over a human nervous system, and it will discover manipulation as a local optimum: infinite scroll, variable-ratio rewards, removed stopping cues, all selected because they move the number, with no person ever sitting down to design a dark pattern. This is Goodhart's law on a human, the same shape as AI misalignment: optimize a proxy hard enough and you get the proxy, not the intent. The real villain is the loss function — which is why the objective-function charter is set in bearings, before the dashboard exists, and audited here, before any metric-driven launch.

This is where the agent era bites:

• The agent treats launch as the end. It ships, the task turns green, and it never wires the feedback loop — no preview to judge perception by eye, no RUM, no error tracking — so the team flies blind exactly where the spec lives (the real user's nervous system).
• The agent instruments by lab, not field, and tracks vanity. It will trust a fast-machine/fast-network lab number, miss the mid-range-phone-on-4G truth, and happily collect every event "for analytics" — monitoring debt and a privacy liability with no decision attached.
• The agent will optimize whatever metric it's handed, including engagement. It feels nothing about the reflective-vs-impulsive-self betrayal, so left alone it points the optimizer at the number that grows, and the optimizer finds the manipulation. The conscience is not a thing the agent supplies — it must be a gate.

Read references/the-membrane.md first — the heart; for lookout, the ethics axis is the spine (friction as the moral primitive, manipulation as bypassing agency, the impulsive-vs-reflective self, A/B as a gradient toward exploitation) along with the mind axis (RUM = psychophysics as numbers). Load at the start, re-check at every gate.

Speak the user's language. Almost every call here is the user's — which metric is the north-star, is this experiment honest, is this default serving the user. Read their fluency and gloss a term on first use (preview deploy, feature flag / progressive rollout, RUM vs lab, Core Web Vitals (LCP/INP/CLS), a guardrail metric, the gradient-descent framing of A/B, the reflective vs impulsive self, friction symmetry / GDPR Art. 7). A metric decision the user can't weigh is an ethics imposed, not shared.