kienbui1995/analyze-chart

Analyze Amplitude Chart

When to Use

• A stakeholder asks "why did this metric drop/spike?"
• You need to understand what is driving a trend in a chart
• A chart shows unexpected behavior and root cause is unknown
• You want to segment a metric to find which user groups are driving a pattern
• Pre-briefing preparation — need to understand the story behind data before presenting
• QA of a chart that may have incorrect filters or event definitions

Core Jobs

1. Pattern Recognition

Load the chart using mcp__Amplitude__query_chart (deep-dive, interactive) or mcp__Amplitude__render_chart (visual snapshot). Identify the primary patterns:

• Trend direction: is the metric rising, falling, or flat over the period?
• Rate of change: is the trend accelerating, decelerating, or linear?
• Magnitude: how large is the change in absolute and relative terms?
• Baseline: where does this metric sit relative to historical norms?

Always extract specific numbers. Avoid vague statements like "the metric went up." Say "DAU increased 23% from 41,200 to 50,700 between March 1 and March 15."

2. Anomaly Identification

Scan for deviations from expected behavior:

• Spikes: sudden upward deviations (check if coincide with releases, campaigns, or bot traffic)
• Drops: sudden downward deviations (check for outages, tracking breaks, platform changes)
• Plateaus: metric stops growing when growth was expected
• Day-of-week effects: lower weekend numbers are often normal, not anomalies
• Incomplete day artifacts: today's number is always lower because the day isn't over

Flag any anomaly with: when it occurred, how large the deviation was, and whether it recovered.

3. Segmentation Analysis

Break the metric down by key dimensions to isolate the driver. Apply segmentation systematically:

• By platform: iOS vs Android vs Web — is the pattern platform-specific?
• By user properties: new vs returning, free vs paid, by geography, by plan type
• By cohort: is the change driven by new users acquired recently or long-term users?
• By feature usage: users who used feature X vs those who didn't

Use mcp__Amplitude__get_event_properties to discover available properties for segmentation. Look for the segment where the pattern is most pronounced — that is where the root cause likely lives.

4. Contextual Investigation

Correlate the finding with external context:

• Product releases: did a deploy happen near the anomaly date?
• Experiments: was an A/B test running that could explain the change?
• Campaigns: did a marketing campaign launch or end?
• External events: holidays, seasonality, competitor activity
• Tracking breaks: did the event volume itself drop (suggesting a logging bug) or did only the metric change?

Ask: "What changed on or just before this date?"

5. Hypothesis Formation

Formulate 2-3 specific, falsifiable hypotheses for what is causing the pattern. Rank by likelihood based on evidence. Each hypothesis should be in the form: "I believe X is happening because I see Y in the data, which would be consistent with Z."

6. Conclusion with Confidence Level

State your conclusion clearly:

• What the chart shows
• The most likely explanation for the pattern
• Your confidence level: High (multiple converging signals), Medium (one strong signal), or Low (insufficient data, hypotheses only)
• What additional investigation would increase confidence

MCP Tools

• mcp__Amplitude__query_chart — deep-dive into a single chart with interactive data
• mcp__Amplitude__render_chart — visual snapshot of a chart for pattern recognition
• mcp__Amplitude__get_event_properties — discover dimensions available for segmentation
• mcp__Amplitude__get_charts — find related charts for cross-referencing
• mcp__Amplitude__query_amplitude_data — run custom queries for additional context

Key Concepts

• Anomaly: A deviation from expected behavior that requires explanation.
• Segmentation: Breaking a metric down by a dimension to isolate which group is driving a pattern.
• Cohort analysis: Comparing user groups defined by when they first performed an action.
• Root cause: The underlying reason a metric changed, as distinct from the surface symptom.
• Correlation vs causation: Two metrics moving together does not mean one causes the other. Distinguish carefully.
• Confidence level: How certain the analysis is, based on the number and quality of converging signals.
• Incomplete day artifact: Today's aggregated number looks lower because not all events have been logged yet. Do not treat this as a drop.

Output Format

Output is written in narrative paragraphs, not bullet lists. The analysis reads like an analyst memo, not a database dump.

Structure:

1. What the chart shows (1 paragraph, specific numbers throughout)
2. Key finding (1-2 paragraphs on the most important pattern or anomaly)
3. Segmentation breakdown (1 paragraph describing which segment drives the pattern)
4. Context and hypotheses (1 paragraph correlating with releases, experiments, or external events)
5. Conclusion (1-2 sentences with confidence level and recommended next step)

Every claim must be backed by a specific number from the data. No vague statements.