Prompt Cards

Goal: analyze onboarding paths with Markov Chains. Data: event sequences for new users first 7 days. Steps: 1) Build state graph of key events; 2) Compute transition probabilities; 3) Identify states increasing probability of {GOAL_EVENT}; 4) Propose nudges. Output: path diagram and recommended guided walkthrough.

Goal: mine internal site search for intent gaps. Data: GA4 view_search_results and search_term params. Steps: 1) Top queries with zero-results; 2) CTR to result pages; 3) Query clusters using embeddings; 4) Map to content backlog. Output: top 20 content/query fixes to capture demand.

Goal: synthesize scroll depth with heatmap insights. Data: GA4 scroll events, heatmap tool exports (e.g., Hotjar). Steps: 1) Join scroll depth to content type/length; 2) Identify false bottoms; 3) Flag modules with low attention but high importance. Output: content redesign recommendations prioritized by impact.

Goal: check experiment readiness and sample size. Data: baseline conversion, traffic, variance. Steps: 1) Minimum detectable effect; 2) Power calculation; 3) Guardrail metrics; 4) Sequential vs fixed-horizon guidance. Output: experiment checklist with sizing and runtime estimates.

Goal: compare attribution approaches. Data: multi-touch journeys (GA4 pathing + ad click logs). Steps: 1) Build rule-based (last/first/linear/time-decay); 2) Build Markov and Shapley using transition matrices; 3) Evaluate stability and business interpretability; 4) Recommend model per objective. Output: attribution comparison deck with channel reallocation.

Goal: model LTV with RFM and Pareto/NBD. Data: orders table + GA4 user_id mapping. Steps: 1) Derive Recency, Frequency, Monetary; 2) Fit Pareto/NBD (BG/NBD) and Gamma-Gamma for monetary; 3) Validate with holdout; 4) Tie back to channels. Output: LTV by segment and channel ROI guidance.

Goal: analyze cohort retention and power-user curve. Data: GA4 users and events, subscription logs if applicable. Steps: 1) Build weekly cohorts by first_visit; 2) Compute W1–W12 retention; 3) Plot power-user curve (events per user per week); 4) Segment by acquisition channel. Output: retention heatmap + target cohort improvements.

Goal: diagnose funnel drop-offs for {GOAL_EVENT}. Data: GA4 BigQuery events with event_params. Steps: 1) Construct funnel steps; 2) Use event timestamps to compute step conversion; 3) Segment by device, traffic source, new/returning; 4) Identify friction fields. Output: funnel report with top 5 friction points and experiment ideas.

Goal: define sessionization robustly and filter bots. Data: raw web logs + GA4 BigQuery. Steps: 1) Build session rules (30m inactivity, campaign change resets); 2) Bot filters using IP ranges, UA regex, request rate z-scores; 3) Compare metrics before/after. Output: reproducible bot filter spec and session ruleset with KPI deltas.

Goal: audit source/medium quality and UTM hygiene for {START_DATE}–{END_DATE}. Data: GA4 BigQuery export `{PROJECT}.{DATASET}.events_*`, ad platform logs, server-side UTMs. Steps: 1) Classify sessions by source/medium/campaign; 2) Flag malformed or missing UTMs; 3) Compute quality metrics (bounce rate, engaged sessions, CVR, AOV, LTV proxy); 4) Rank sources by quality vs spend. Output: table of sources with UTM fix checklist and top 5 cleanup priorities.

Goal: nowcast GDP and inflation using alternative data. Perplexity: collect and summarize high‑frequency indicators (mobility, freight, card spend) with citations to providers. Bloomberg: ECST<GO> and BQL pulls for weekly/monthly proxies; GP<GO> to chart; ECO<GO> for official releases to benchmark. Deliverable: nowcast vs consensus gap and trade implications.

Goal: interpret China A‑shares sentiment and policy. Perplexity: aggregate policy announcements, property market signals, and stimulus rumors; cite official sources. Bloomberg: CN<GO>/NI<GO> for policy headlines; ECST<GO> for credit/PMI series; HP<GO> for CSI300 sector moves; RV<GO> for valuation reset. Deliverable: sentiment scorecard and sector positioning notes.