evylegzhanin/agent-with-sandbox/eda

Exploratory Data Analysis (EDA)

When this skill applies, treat the task as requiring a disciplined exploration pass before conclusions or heavy modeling.

Load python-scripts via read_skill if you have not already; use code in every write_python_file call, then run_python_file.

Protocol

1. Inventory inputs — if the harness injected an Automatic input profile, use exact sheet and column names from it. Otherwise call list_input_files, then staged Python to list sheets or columns when needed.
2. Profile each relevant table — use pandas directly (df.shape, df.dtypes, df.isnull().sum(), df.describe(), df.head()). Example:

```python import pandas as pd

df = pd.read_excel("/workspace/input/sales.xlsx", sheet_name="Sales") print("shape:", df.shape) print(df.dtypes) print(df.isnull().sum()) print(df.describe(include="all")) df.to_csv("/workspace/out/sales.csv", index=False) ``` Parquet is not supported in the sandbox; use CSV for intermediates.

1. Check structure — detect index columns, duplicate keys, obvious ID columns, date/time fields, mixed-type columns, and encoding issues.
2. Distributions and relationships — where useful, print or plot univariate summaries (histograms, KDE, boxplots) and simple bivariate views (scatter for numeric pairs, grouped counts). Prefer a few high-signal visuals over many redundant charts. Save plots under /workspace/out.
3. Outliers and data quality — flag extreme values, impossible ranges, or inconsistent categories; quantify how many rows are affected before proposing fixes.
4. Communicate findings — summarize what you learned in stdout (and in the final user-facing answer) before task-specific transformations or exports.

Reuse files under /workspace/out across staged Python runs when building follow-on plots or decks. Follow the [Skill: python-scripts] checklist for script shape and error recovery.