richardhe-fundamenta/practical-gcp-examples/data-analysis

data-analysis

Use this skill when the user wants to analyze data and/or get a chart or report — especially when they uploaded a data file.

Steps:

1. Load the uploaded data by filename (it's in the working directory), e.g.

df = pandas.read_csv("data.csv") (or read_excel). Inspect columns/dtypes as needed.

1. Compute the requested aggregation/analysis with pandas (group-by, resample, etc.).
2. Produce output files — they are returned to the user as attachments:

• Chart → save an image: matplotlib/seaborn fig.savefig("chart.png", dpi=150, bbox_inches="tight").
• Tabular report → write report.xlsx (df.to_excel) or report.csv.
• Do not print raw bytes or a giant table to stdout — write a file instead.

1. Print a short text summary (the headline finding) to stdout alongside the file(s).
2. Call run_code with the full script.

Displaying charts in the UI:

• After run_code returns, any image you saved is reported back as a short placeholder token,

e.g. chart.png -> {{chart:ab12}}. To show it, make ONE send_a2ui_json_to_client call whose a2ui_json is a JSON list with BOTH a `beginRendering` and a `surfaceUpdate` — never beginRendering alone (that's an empty surface). Lay the surface out top-to-bottom: a title Text, a `thinking` Text (2–4 short sentences explaining how you approached the task and what the data shows — your reasoning, in plain language), then the chart in an Image whose url is the token copied EXACTLY (it resolves to the real image). Column.children uses {"explicitList": [<ids>]}: ``json [ {"beginRendering": {"surfaceId": "report", "root": "root", "styles": {}}}, {"surfaceUpdate": {"surfaceId": "report", "components": [ {"id": "root", "component": {"Column": {"children": {"explicitList": ["title", "thinking", "chart"]}}}}, {"id": "title", "component": {"Text": {"text": {"literalString": "<headline>"}, "usageHint": "h3"}}}, {"id": "thinking", "component": {"Text": {"text": {"literalString": "<2-4 sentences: how you analyzed it + what the data shows>"}}}}, {"id": "chart", "component": {"Image": {"url": {"literalString": "{{chart:ab12}}"}}}} ]}} ] ``

Guidance:

• Available packages include pandas, numpy, polars, pyarrow, duckdb, scipy, scikit-learn,

statsmodels, matplotlib, seaborn, altair, plotly, openpyxl.

• Prefer matplotlib savefig("chart.png", ...) so the user gets a rendered image.
• One clear chart per request unless asked otherwise; label axes and give it a title.
• No network access; work only from the uploaded data.
• Never hardcode, reconstruct, or invent the dataset. Analyze only the actual file in the

working directory (open it by name). If the file isn't there, do NOT recreate it from memory — ask the user to attach it and stop.

• If run_code returns an error, fix the script and retry.