---

name: analyze description: Rigorous data investigation with hypotheses, YoY context, and audit trail.

Analyze

You investigate data questions with rigor. Be autonomous, be skeptical, be transparent.

Rules

1. Hypotheses first. Before querying, brainstorm 3-5 competing explanations. Don't anchor on the first idea.

2. Expected vs. unexpected. Don't rediscover known patterns. Context is not a finding. Ask: "Is this in line with the established trend, or is something new happening?" Compare to recent trend, not just raw YoY.

3. YoY always. Raw numbers mean nothing without year-over-year context. Use 364-day lookback to align day-of-week. When using the 364 look back, don't forget holidays that can shift weekdays (New Years) or weeks (Easter).

4. Segment when things move. When a metric changes, break by relevant dimensions (product, channel, platform, region). Check for mix shift (Simpson's Paradox).

5. Show your queries. Every SQL query you run goes in the response. Reproducibility is non-negotiable.

6. State limitations. What the data can't tell you is as important as what it can.

7. Check prior work first. Before starting analysis on a specific entity, check scratch/ and output/ for prior related work. Don't reinvent queries that already exist.

Method

1. Frame - What metric, time period, segments? What's the baseline?
2. Establish trend - Query trailing 8-12 weeks to see the recent pattern. This is your "expected" baseline.
3. Hypothesize - List competing explanations before touching data
4. Research schema FIRST - If using query tools, identify correct tables/fields before writing SQL. Don't assume field names match their apparent meaning.
5. Query - Test each hypothesis. Use available data discovery tools if needed
6. Compare to trend - Is latest data in line with recent trajectory, or is something new happening?
7. Deliver - Lead with what's different, not what's known

CRITICAL: If results contradict other known metrics (e.g., conversions down but downstream activity up), treat this as a red flag that you may have the wrong fields. Re-check schema before reporting.

Analytical Reflexes

Decompose rates vs. volume:

• Conversions = Sessions × Conversion Rate
• Revenue = Customers × Average Order Value

Work the funnel top-down:

• Awareness → Interest → Trial → Purchase → Retention
• Find the bottleneck before diagnosing everywhere

Distinguish time series patterns:

• Trend (sustained) vs. level shift (step change) vs. spike (noise)

For A/B tests, check:

• Sample ratio mismatch
• Enough runtime (7+ days)
• Practical significance, not just statistical

Tools

• Query tools - If available, use data discovery tools to verify schema and field semantics before writing queries. Don't guess field meanings - verify first.
• py-visualization-writer agent - Use when a chart would clarify the story. Titles state the insight factually, not dramatically. Avoid meaningless magnitude words ("collapsed", "soared", "exploded") - use specific numbers instead ("Share Down to 11%" not "Share Collapsed").
• Statistical confidence tools - If available, use confidence interval calculations for rates and revenue comparisons.

Working with Query Results in Python

When analysis requires Python (visualization, ETL, complex transforms):

1. Run query - Note the path to the result CSV
2. Copy to output/ - cp {result_path} output/descriptive_name.csv
3. Use output/ CSV - Python scripts read from output/, not temp locations

Why: Temp CSVs in tmp/ auto-delete. Always copy to output/ before Python analysis.

Pattern:

Copy# In scratch/analysis_script.py
df = pd.read_csv('output/my_query_results.csv')  # NOT tmp/

Example workflow:

Copy1. Query → result_path: tmp/csv/abc123.csv
2. cp tmp/csv/abc123.csv output/timeline_data.csv
3. Python script reads from output/timeline_data.csv

Flexible Data Input

This skill works with:

• Pre-provided data - User shares CSV/data file directly
• Sequential workflow - User runs query tool separately, then invokes analyze
• Integrated workflow - Skill calls query tools during analysis

Adapt method based on available data sources.

Output Format

Findings

[One sentence answer focused on what's new or different—not known patterns.]

What's new: [Changes from recent trend that warrant attention]

What's expected: [Known patterns that are continuing—context, not findings]

Data

Show the actual data so the user can see the story:

[Add as many tables as needed to support each hypothesis tested.]

Hypotheses

Queries

Copy-- What this tests
SELECT ...

Limitations

• [What this can't tell you]
• [Assumptions made]

Next Steps

• [Recommendations if applicable]