data-analysis（数据分析）

Data Analysis

When to use this skill

• The user has a dataset, export, report extract, query result, or shaped event / telemetry table and wants evidence-backed conclusions.
• The task is to understand what changed, compare segments, summarize performance, or explain anomalies in business terms.
• The request mentions CSV, JSON, SQL tables, retention, cohorts, funnels, conversion, spend, telemetry, event exports, or KPIs.
• The work needs data-quality checks before conclusions.
• The user needs a concise analysis narrative, not just raw code snippets.

Do not use this skill as the main workflow when:

• The main goal is repeated anomaly or code-pattern scanning across code/data assets → use pattern-detection.
• The main goal is building or tuning a specific BI dashboard / Looker Studio + BigQuery workflow → use looker-studio-bigquery.
• The task is repository navigation or call-site tracing rather than dataset reasoning → use codebase-search.
• The problem is raw log triage / incident reconstruction rather than dataset analysis → use log-analysis.

Core idea

Data analysis is a staged reasoning workflow:

1. clarify the decision question
2. profile the data and trust level
3. choose the cheapest analysis lane that can answer it
4. separate observation from interpretation
5. finish with evidence, caveats, and next actions

Do not jump straight into charts or code. The goal is decision-quality analysis.

Instructions

Step 1: Frame the analysis question

Before touching the data, define:

• Decision to support — what action or judgment depends on this analysis?
• Primary metric(s) — conversion, retention, revenue, latency, churn, balance, spend efficiency, etc.
• Dimensions / segments — time, channel, cohort, region, plan, device, feature flag, player segment
• Comparison mode — before/after, control/treatment, top vs bottom segments, expected vs actual
• Time window — day/week/month/release/experiment period

If the request is vague, restate it as:

"We need to explain [metric/outcome] for [audience] over [time window] and identify the strongest drivers or caveats."

Step 2: Run a trust check before analysis

Always start with data-quality triage.

Minimum trust checklist

• row count / extract size
• schema and types
• missing values / null-heavy columns
• duplicates or repeated IDs
• time range coverage and timezone assumptions
• segment completeness (channels, countries, devices, builds, player groups)
• obvious join / aggregation errors
• outliers or impossible values

Default check pattern:

import pandas as pd

# df = pd.read_csv(...)
print(df.shape)
print(df.dtypes)
print(df.head())
print(df.isna().sum().sort_values(ascending=False).head(15))
print(df.duplicated().sum())

If trust is low, stop promising conclusions and explicitly switch the output to:

• what is trustworthy
• what is suspect
• what additional cleanup or data is needed

Step 3: Choose the analysis lane

Pick the cheapest lane that can answer the question. Escalate only when needed.

Step 4: Use the right analysis pattern

Pattern A — Change explanation

Use for: experiments, release effects, KPI jumps/drops, spend shifts, gameplay balance changes.

Checklist:

1. define baseline and comparison window
2. confirm denominator / assignment integrity when this is an experiment or rollout comparison
3. compute absolute + relative deltas
4. break the change by top segments or drivers
5. test whether the change is broad or concentrated
6. call out confounders (seasonality, launch, tracking changes, sample size, significance/confidence limits)

Pattern B — Segment comparison

Use for: channel quality, user tiers, device classes, regions, player cohorts.

Checklist:

1. rank segments by the primary metric
2. include sample size / denominator
3. compare both rate and volume
4. watch for Simpson's-paradox-style aggregation traps
5. explain what likely differentiates top vs bottom groups

Pattern C — Funnel / retention analysis

Use for: signup, purchase, onboarding, feature adoption, live-ops progression.

Checklist:

1. define each stage/event clearly
2. compute stage counts and conversion/drop-off rates
3. segment by acquisition source, cohort, platform, build, or player type
4. identify the highest-leverage drop-off point
5. distinguish instrumentation gaps from genuine behavior problems

Pattern D — Telemetry / event analysis

Use for: gameplay telemetry, product event streams, operational exports.

Checklist:

1. map raw events to derived metrics
2. group by session/build/feature/segment/time
3. identify spikes, sinkholes, and suspicious clusters
4. separate normal variation from suspicious outliers
5. route sustained anomaly-hunting work to pattern-detection if the task becomes detection-first

Step 5: Keep observations separate from interpretation

Structure findings in three layers:

1. Observation — what the data literally shows
2. Interpretation — likely meaning or driver
3. Caveat / confidence — what could weaken the conclusion

Good example:

• Observation: conversion dropped 6.2% week-over-week, concentrated in mobile Safari traffic.
• Interpretation: the decline is likely connected to the recent checkout UI change on smaller screens.
• Caveat: tracking for one payment method was also modified that week, so attribution is medium confidence.

Step 6: Return a decision-ready output

Default output shape:

## Analysis brief
- Goal: [decision question]
- Data source: [files / tables / export scope]
- Trust level: high | medium | low
- Lane used: spreadsheet triage | SQL slicing | notebook/statistical | summary-only

## Key findings
1. [finding]
2. [finding]
3. [finding]

## Supporting evidence
- [metric / segment / comparison]
- [metric / segment / comparison]

## Caveats
- [missing data / sample bias / instrumentation / seasonality]

## Recommended next actions
- [decision / follow-up slice / dashboard handoff / instrumentation fix]

If the user asked for recommendations, tie each recommendation to a specific finding. If the user only asked for analysis, stop at evidence + caveats.

Step 7: Route out when analysis stops being the bottleneck

Hand off when the next step is a different job:

• Repeated anomaly hunting or rule-based scanning → pattern-detection
• Dashboard construction / BigQuery-connected reporting → looker-studio-bigquery
• Raw log triage before dataset shaping → log-analysis
• Repo/code investigation to find instrumentation or metric definitions → codebase-search

Examples

Example 1: Experiment analysis

Prompt:

Analyze this CSV export and tell me what changed after the pricing experiment.

Good response shape:

• define baseline vs experiment window
• check data coverage and segment completeness
• report overall delta plus segment breakdown
• identify strongest likely drivers and caveats

Example 2: Marketing + product analysis

Prompt:

We have app event logs and marketing spend by channel; find the main retention and CAC patterns.

Good response shape:

• separate acquisition and retention metrics
• compare rate and volume by channel/cohort
• note trust limits if joins or attribution windows are unclear
• summarize high-leverage channel differences

Example 3: Game telemetry analysis

Prompt:

Review this gameplay telemetry extract and summarize balance issues and suspicious outliers.

Good response shape:

• map events to gameplay metrics
• compare player/build/weapon/level segments
• separate broad balance patterns from suspicious outliers
• route repeated anomaly detection to pattern-detection if needed

Example 4: PM / ops export triage

Prompt:

I exported a dashboard to CSV; help me explain the KPI drop for leadership.

Good response shape:

• start with trust checks on the export
• identify the metric, time window, and comparison baseline
• produce a concise leadership-ready memo with evidence and caveats

Best practices

1. Start from the decision question, not the chart type.
2. Run data-quality checks before interpretation.
3. Always include sample size / denominator context when comparing segments.
4. Prefer the cheapest sufficient lane instead of defaulting to heavy notebooks.
5. Separate observation, interpretation, and caveat so the analysis stays honest.
6. Route dashboard-building and anomaly-detection work to adjacent specialist skills when they become the real task.

References

• Project Jupyter
• Pandas getting started tutorials
• DuckDB Jupyter guide
• GA4 share & export reports

Output format

Use a brief, findings-first summary with trust level, key evidence, caveats, and explicit next actions or handoffs.