cja-kpi-pulse

KPI Pulse (Customer Journey Analytics)

Produce a compact KPI digest in under 2 minutes. The goal is a crisp answer to "how did we do?" — not a deep-dive, not a data dump. Each KPI gets a scorecard showing current value, period-over-period change, trend direction, and the top dimension breakdown that explains any movement.

CJA MCP Tools Used

• describeCja(DATAVIEW_CONTEXT_GUIDE) — understand the data view context
• listComponentUsage — find the most-used metrics (the org's real KPIs)
• findMetrics — resolve metric IDs from user-specified names
• findCalculatedMetrics — include custom KPIs if present
• runReport — pull metric values for current and prior periods
• searchDimensionItems — top dimension breakdown for movers

Phase 0 — Setup

1. Call findDataViews to list available data views.
2. If the user hasn't specified a data view, present the list and ask which to use.
3. Call setDefaultSessionDataViewId with the chosen ID.
4. Call describeCja("DATAVIEW_CONTEXT_GUIDE") to load data view context.

Record the data view's first-day-of-week as WEEK_START_DOW and timezone as TIMEZONE. If the context guide does not return a week-start value, default to Monday (ISO 8601). You will use both in Phase 1.1.

1. Clarify the monitoring scope: which KPIs to track and the comparison period (e.g., WoW, MoM, vs. target).

Phase 1 — Clarify Scope

1.1 Determine the reporting period

If the user did not specify a period, ask one question:

"What time window would you like? Options: last 7 days, last 30 days, this week vs last week, this month vs last month, or a custom range."

Default to this week vs last week if no answer is given.

Map the answer to two date ranges:

• Period A (current): e.g., "thisWeek", "thisMonth", last 7 days
• Period B (comparison): e.g., "lastWeek", "lastMonth", prior 7 days

Calendar rule (mandatory):

Use WEEK_START_DOW from Phase 0 to define what "week" means. The current period (Period A) and the comparison period (Period B) MUST use the same first-day-of-week — i.e., both periods' startDate fall on the same day-of-week, both are exactly equal length, and the comparison period ends immediately before the current period starts. Never mix conventions (e.g., a Mon–Sun current with a Sun–Sat prior) within the same pulse run. Pick the boundary once, then derive both periods from it. For custom date ranges, compute Period B as the equal-length window ending immediately before Period A starts.

Sanity check before calling `runReport`: confirm periodA.startDate and periodB.startDate are the same day-of-week and that periodA.startDate - periodB.endDate == 1 day. If not, recompute.

1.2 Determine the metrics

If the user named specific metrics, resolve them with findMetrics or findCalculatedMetrics. Otherwise, discover the top 5–8 KPIs automatically:

listComponentUsage(componentType: "metric")
listComponentUsage(componentType: "calculatedMetric")

Note: listComponentUsage may return an empty list for data views with no usage history. If it returns empty, fall back to:

findMetrics(searchQuery: "sessions visits revenue orders")
findMetrics(searchQuery: "page views cart conversion")

Pick the most business-relevant metrics from the results (sessions, orders, revenue, product views, cart views, people — in that priority order).

Deduplicate: if a built-in metric and a calculated metric measure the same thing, keep only the calculated metric (it's more intentional).

Final list: 5–8 metrics. More than 8 KPIs in a pulse report is noise.

Phase 2 — Pull Current and Prior Period Data

Run a single runReport call per period with all KPI metrics included. Use one call for Period A and one for Period B to minimize round-trips. Use a summary dimension (e.g., variables/daterangeday) and limit: 1 to get aggregate totals from summaryData.totals in the response.

runReport(
  dimensionIds: "variables/daterangeday",
  metricIds: "metrics/visits,metrics/visitors,metrics/orders_1_1,metrics/productListItems.priceTotal,metrics/cart_views",
  startDate: "<periodA start>T00:00:00",
  endDate: "<periodA end>T23:59:59",
  page: 0,
  limit: 1
)

runReport(
  dimensionIds: "variables/daterangeday",
  metricIds: "metrics/visits,metrics/visitors,metrics/orders_1_1,metrics/productListItems.priceTotal,metrics/cart_views",
  startDate: "<periodB start>T00:00:00",
  endDate: "<periodB end>T23:59:59",
  page: 0,
  limit: 1
)

Read aggregate totals from summaryData.totals (not row data), which gives you the full-period sum for each metric in the order they were listed.

Capture for each metric:

• valueA (current period)
• valueB (comparison period)
• delta = valueA − valueB
• pctChange = (delta / valueB) × 100, rounded to 1 decimal

Phase 3 — Classify Trends

For each KPI, assign a trend indicator:

• ↑ Up if pctChange > +3%
• ↓ Down if pctChange < −3%
• → Flat if −3% ≤ pctChange ≤ +3%

Assign a signal color:

• For "higher is better" metrics: ↑ = green, ↓ = red, → = grey
• For "lower is better" metrics (bounce rate, error rate): ↑ = red, ↓ = green

Phase 4 — Top Mover Drill-Down

For the 1–2 metrics with the largest absolute % change, find what's driving the movement. Run a dimension breakdown for the current period:

runReport(
  dimensionIds: "variables/marketing_channel",
  metricIds: "<moving metric id>",
  startDate: "<periodA start>T00:00:00",
  endDate: "<periodA end>T23:59:59",
  page: 0,
  limit: 5
)

Note: Use variables/marketing_channel (not variables/marketingchannel) — verify the exact dimension ID with findDimensions(searchQuery: "marketing channel") if unsure.

Compare dimension values between Period A and Period B to identify the top contributor to the change. This becomes the "What drove it" entry in the report.

Phase 5 — Generate HTML Report

Generate the KPI Pulse HTML report INLINE — do not use a Python script. Build the HTML string directly from the collected data and output it as a code block the user can save, or write it to /tmp/cja_kpi_pulse_report_<YYYY-MM-DD_HHMMSS>.html using a one-line bash command.

Rendering rules — apply consistently across runs

Two runs of this skill on the same data view + period must render identically (modulo the generation timestamp). The rules below pin the formatting choices that the AI would otherwise drift on.

Number formatting

• KPI values (the big number in each tile) — use full digits with

thousands separators (8,160, 77,584, 1,250,000). Do NOT use SI suffixes like K or M, even for large values. Executives want exact numbers, not abbreviations.

• Percent change (in pills and narrative bullets) — always one decimal

place, rounded half-away-from-zero. For example, −23.55% displays as −23.6%, never −23.5%. Compute on full-precision values; round only at display time.

• Percentage-point change (for already-percentage metrics like Conversion

Rate or Bounce Rate) — same rounding, suffix pp. Example: +0.40 pp.

• Currency — $ prefix with thousands separators and no decimals for

values ≥ $100 ($1,240,000); cents only when value < $100 ($45.20).

Null / missing data handling

A KPI tile must reflect what the data view actually returned. The AI must not silently substitute a different metric or hide a tile to make the report look cleaner.

• Both periods return 0 or NULL for a KPI being rendered: render the tile

with kpi-value = Data unavailable, pill class flat, pill text ⚠ N/A, and prior text = Both periods returned no data — validate instrumentation. The tile stays in the grid; do not omit it.

• One period returns valid data, the other 0 / NULL: render the tile with

the valid value as kpi-value, pill class flat, pill text ⚠ N/A, and prior text = Prior {period_noun}: no data.

• Never substitute a derived metric (e.g., adding "Conversion Rate"

because Revenue came back $0). The visible KPI set MUST match the metrics selected for this run.

HTML Template

Read `template.html` and use it verbatim. Do not improvise the HTML structure or CSS — only fill in the {PLACEHOLDER} tokens ({ORG_NAME}, {PERIOD_LABEL}, {COMPARISON_LABEL}, {DATA_VIEW}, {GENERATED_DATE}, {METRIC_NAME}, {FORMATTED_VALUE_A}, {FORMATTED_VALUE_B}, {PCT_CHANGE}, {VALUE_A}, {VALUE_B}, {DELTA}, {ARROW}) and repeat the KPI tile / detail row / mover row blocks once per data item. Preserve the .up | .down | .flat and .green | .red | .yellow | .grey modifier classes per the trend rules in Phase 3.

Phase 6 — Deliver the Report

After generating the HTML:

1. Write it to /tmp/cja_kpi_pulse_report_<YYYY-MM-DD_HHMMSS>.html
2. Open with open /tmp/cja_kpi_pulse_report_<YYYY-MM-DD_HHMMSS>.html
3. Provide a 3–5 line text summary inline in the chat:

KPI Pulse — This Week vs Last Week

↑ Revenue: $1.24M (+8.2%)  — Paid Search drove most of the gain
↓ Conversion Rate: 2.1% (−0.4pp) — Drop in mobile checkout
→ Sessions: 540K (+1.1%)  — Flat week-over-week
↑ Orders: 11,340 (+6.7%)  — Product page improvements appear to be working
↓ Bounce Rate: 43.2% (+2.1pp) — Worth monitoring next week

The text summary gives immediate value even without opening the HTML file.

Important Guardrails

• Read-only monitoring. Never modify metrics, segments, or projects.
• Use consistent date ranges. Week-over-week and month-over-month comparisons must use equal-length periods.
• Flag anomalies, don't diagnose them. The pulse report surfaces significant deviations — deep root cause analysis belongs in the anomaly triage skill.
• Respect business calendar. Holiday periods, campaigns, and seasonal patterns affect normal variance — note context when flagging anomalies.
• Cap metric count. Monitor up to 10–15 KPIs per pulse; more than that dilutes focus. Ask the user to prioritize if they specify too many.
• Note data freshness. If the most recent data point is older than expected, warn the user before presenting the pulse.

Example Interaction

"Give me a quick pulse on our key metrics for this week."

1. Setup: Confirm data view with findDataViews. User selects their main data view. Call setDefaultSessionDataViewId.
2. Scope: Ask "Which KPIs should I include?" User says: "Sessions, Revenue, Conversion Rate, and Average Order Value."
3. Data pull: Run runReport for current week vs. prior week for all four metrics.
4. Analysis: Sessions +8% WoW (within normal range). Revenue +3% WoW. Conversion Rate -12% WoW — flagged as anomalous. AOV +17% WoW — notable positive.
5. Summary: Present a KPI scorecard with traffic-light status (green/yellow/red), highlight the Conversion Rate drop as needing investigation, and note that the AOV increase partially offsets it.

Error Handling

• If runReport returns no data for Period B (comparison is too far in the

past or data view lacks history), show "N/A" for the delta and flag it with a grey badge.

• If a metric returns null, display "—" rather than 0 to avoid false

impressions of zero performance.

• If fewer than 3 metrics are available, warn the user that the pulse may be

incomplete and suggest they verify the data view is correctly configured.