richardhe-fundamenta/practical-gcp-examples/analyst-chart-table

Analyst chart + table

This skill turns a small set of already-validated aggregates into a single secure visual: one chart with the finding stated in the title, and a compact table of the exact numbers beneath it.

Inputs you receive

• A JSON object of aggregated rows (small — already reduced by SQL upstream).
• The business question being answered.
• Optionally: a target/prior-period value for delta context.

Do not query data here. Data arrives pre-aggregated and pre-validated from the harness. Your job is interpretation + rendering only.

Procedure

1. Pick the chart type from the question shape:

• trend over time -> line
• comparison across categories -> bars
• if neither fits cleanly, default to bars and say so in the title.

1. Derive the headline finding — the single most decision-relevant fact

(biggest mover, threshold crossed, outlier segment). This becomes the title, phrased as a conclusion ("APAC churn doubled in Q1"), never a label ("Churn by region").

1. Show change, not just level — annotate the delta vs prior period or a

reference line. Analysts think in deltas.

1. Generate Python rendering code (see Execution). Pass data as JSON; never

build markup by string-concatenating data values.

1. Attach the compact table of exact numbers beneath the chart.
2. Write one plain-English "so what" line.

Execution

You (the model) generate Python rendering code; the harness runs it in an isolated sandbox (no network, no credentials) via the render_chart tool.

Data source: The harness writes the rows from the most recent successful validated query to data.json in the working directory, as:

{ "rows": [ { "col": value, ... }, ... ] }

rows is a list of record dicts straight from BigQuery — these are the ONLY data you may chart. Your code reads data.json, shapes/aggregates rows (e.g. pivot a month/category/value layout into series), and renders. Never hardcode values and never invent labels or numbers not present in rows; you cannot pass data yourself, which guarantees the chart reflects real queried data.

Packages: Use ONLY preinstalled packages listed in references/available-packages.md. Never pip install; no network access is available in the sandbox.

Output: Emit exactly ONE PNG written to output.png. Use matplotlib with the Agg backend (matplotlib.use("Agg")). Save with fig.savefig("output.png", bbox_inches="tight").

Style reference: references/example_render.py shows the intended look and shape — palette (#b07a3c, #5f817b, etc.), headline-as-title (bold, left-aligned, conclusion phrased), compact table beneath with dark header row, and a "so what" line in italic accent colour. Adapt it; do not assume it is present in the sandbox or call it directly.

Safety rules for generated code:

• Do not build markup by string-concatenating data values; treat data as data.
• No import subprocess, import os.system, network calls, or file writes

other than output.png.

• The PNG default is secure: no JavaScript in the artifact, portable to

email/PDF/chat.

• Only generate the interactive-HTML variant when hover/zoom genuinely helps AND

the harness confirms a trusted display surface. The harness — not this skill — applies iframe sandboxing and CSP to any HTML output. See references/output-contract.md and references/security-notes.md.

Output rule

Exactly one takeaway, stated in words, on one visual. If the data supports several findings, pick the most decision-relevant one; do not add panels.