Chart Deep Dive
When to Use
- A metric spiked or dropped unexpectedly
- You need to understand what’s driving a trend
- Preparing a detailed, evidence-backed analysis for stakeholders
- Investigating differences between user or event segments
Instructions
Step 0: Identify the Chart
- Accept a chart URL or chart ID
- If the user provides a URL, use
Amplitude:get_from_url to extract the chart ID
- If no chart identifier is provided, ask explicitly for the chart URL or ID and stop
Step 1: Retrieve and Validate Chart Data (Mandatory)
- Use Reading chart data to retrieve the chart definition and data
- If chart data cannot be retrieved or is empty, do not proceed
- Explain what’s missing (time range, event, filters, permissions)
- Ask the user to correct the chart or provide a valid chart
Capture and restate:
- Metric being measured
- Time range and granularity
- Chart type (e.g. time series, funnel, retention)
- Existing filters, segments, or breakdowns
Step 2: Identify the Pattern and Change Window
Use Analyzing chart to characterize what’s happening:
- Spike / Drop: Sudden change on specific date(s)
- Trend: Gradual increase or decrease over time
- Seasonality: Recurring weekly or monthly patterns
- Anomaly: Deviation from recent baseline or historical behavior
Explicitly identify:
- The window of change (start/end)
- Direction and magnitude of the change
- Baseline period used for comparison (default: previous equal-length period)
Step 3: Investigate Likely Drivers (Bounded)
Instead of broad slicing, use guided segmentation:
- Use Finding the right event properties to identify the most relevant properties for explaining the change
- Select up to 9 high-signal properties (e.g. platform, country, plan, version)
- Re-run Analyzing chart with these properties in mind to determine:
- Which segments contribute most to the change
- Whether the pattern is localized or broad-based
- Only fetch up to 3 charts at a time when using
Amplitude:query_charts
Avoid testing more than 9 properties in aggregate unless the user explicitly asks for deeper exploration.
Step 4: Correlate with Context (Required for Anomalies)
For spikes, drops, or unexpected shifts, gather contextual signals in the same timeframe:
- Use Getting experiments to identify active experiments or flags
- Use Getting deployments to identify releases or rollouts
- Use Searching for content to surface annotations or relevant documentation
- Use
Amplitude:get_feedback_insights to search customer feedback trends that might explain the change
- Use
Amplitude:get_feedback_mentions to pull in specific customer mentions if there's a likely feedback trend tied to what's being explained.
Determine whether any contextual changes align temporally with the chart pattern.
Step 5: Synthesize Findings
Present a structured, decision-ready analysis:
What Happened
Clear description of the observed pattern and magnitude
When
Exact timeframe and comparison baseline
Primary Hypothesis
Most likely explanation based on chart data and contextual signals
Supporting Evidence
- Key metrics
- Segment contributions
- Relevant experiments, deployments, or annotations
Alternative Explanations
1–3 plausible alternatives and why they are less likely
Impact
Quantify impact where possible (users, events, conversion, revenue proxy)
Recommended Next Step
One clear follow-up action (e.g. deeper segment, experiment review, instrumentation check)
Always include:
- Chart name
- Chart ID
- Link back to the chart
- Coverage (e.g. properties tested, segments analyzed)
Best Practices
- Always compare against a clear baseline period
- Distinguish observations from hypotheses
- Prefer high-signal segmentation over exhaustive slicing
- Note data quality issues (low volume, incomplete periods, heavy “(none)” values)
- Do not create or edit charts unless the user explicitly asks
1---
2name: analyze-chart
3description: Performs deep analysis of a specific Amplitude chart to explain trends, anomalies, and likely drivers. Use when a metric looks unusual, investigating a spike or drop, or understanding the "why" behind numbers.
4x-amp-exclude-when-flags: [mcp-consolidate-charts]
5metadata:
6 source: https://skillmd.com/skills/amplitude/analyze-chart
7---
8
9# Chart Deep Dive
10
11## When to Use
12
13- A metric spiked or dropped unexpectedly
14- You need to understand what’s driving a trend
15- Preparing a detailed, evidence-backed analysis for stakeholders
16- Investigating differences between user or event segments
17
18## Instructions
19
20### Step 0: Identify the Chart
21
22- Accept a chart **URL or chart ID**
23- If the user provides a URL, use `Amplitude:get_from_url` to extract the chart ID
24- If no chart identifier is provided, ask explicitly for the chart URL or ID and stop
25
26---
27
28### Step 1: Retrieve and Validate Chart Data (Mandatory)
29
30- Use **Reading chart data** to retrieve the chart definition and data
31- If chart data cannot be retrieved or is empty, **do not proceed**
32 - Explain what’s missing (time range, event, filters, permissions)
33 - Ask the user to correct the chart or provide a valid chart
34
35Capture and restate:
36- Metric being measured
37- Time range and granularity
38- Chart type (e.g. time series, funnel, retention)
39- Existing filters, segments, or breakdowns
40
41---
42
43### Step 2: Identify the Pattern and Change Window
44
45Use **Analyzing chart** to characterize what’s happening:
46
47- **Spike / Drop**: Sudden change on specific date(s)
48- **Trend**: Gradual increase or decrease over time
49- **Seasonality**: Recurring weekly or monthly patterns
50- **Anomaly**: Deviation from recent baseline or historical behavior
51
52Explicitly identify:
53- The **window of change** (start/end)
54- Direction and magnitude of the change
55- Baseline period used for comparison (default: previous equal-length period)
56
57---
58
59### Step 3: Investigate Likely Drivers (Bounded)
60
61Instead of broad slicing, use **guided segmentation**:
62
631. Use **Finding the right event properties** to identify the most relevant properties for explaining the change
642. Select **up to 9 high-signal properties** (e.g. platform, country, plan, version)
653. Re-run **Analyzing chart** with these properties in mind to determine:
66 - Which segments contribute most to the change
67 - Whether the pattern is localized or broad-based
68 - Only fetch up to 3 charts at a time when using `Amplitude:query_charts`
69
70Avoid testing more than 9 properties in aggregate unless the user explicitly asks for deeper exploration.
71
72---
73
74### Step 4: Correlate with Context (Required for Anomalies)
75
76For spikes, drops, or unexpected shifts, gather contextual signals in the same timeframe:
77
78- Use **Getting experiments** to identify active experiments or flags
79- Use **Getting deployments** to identify releases or rollouts
80- Use **Searching for content** to surface annotations or relevant documentation
81- Use `Amplitude:get_feedback_insights` to search customer feedback trends that might explain the change
82- Use `Amplitude:get_feedback_mentions` to pull in specific customer mentions if there's a likely feedback trend tied to what's being explained.
83
84Determine whether any contextual changes align temporally with the chart pattern.
85
86---
87
88### Step 5: Synthesize Findings
89
90Present a structured, decision-ready analysis:
91
921. **What Happened**
93 Clear description of the observed pattern and magnitude
94
952. **When**
96 Exact timeframe and comparison baseline
97
983. **Primary Hypothesis**
99 Most likely explanation based on chart data and contextual signals
100
1014. **Supporting Evidence**
102 - Key metrics
103 - Segment contributions
104 - Relevant experiments, deployments, or annotations
105
1065. **Alternative Explanations**
107 1–3 plausible alternatives and why they are less likely
108
1096. **Impact**
110 Quantify impact where possible (users, events, conversion, revenue proxy)
111
1127. **Recommended Next Step**
113 One clear follow-up action (e.g. deeper segment, experiment review, instrumentation check)
114
115Always include:
116- Chart name
117- Chart ID
118- Link back to the chart
119- Coverage (e.g. properties tested, segments analyzed)
120
121---
122
123## Best Practices
124
125- Always compare against a clear baseline period
126- Distinguish **observations** from **hypotheses**
127- Prefer high-signal segmentation over exhaustive slicing
128- Note data quality issues (low volume, incomplete periods, heavy “(none)” values)
129- Do **not** create or edit charts unless the user explicitly asks