Agent Skills: Python Data

Python Data

Load python3-core for standing defaults. Load python3-typing for boundary schemas. Load python3-testing for parser and edge-case tests.

Quality Checklist

• [ ] Schema validated at first stable ingress point — not deep in transforms
• [ ] dtype= explicit in pd.read_csv() / pd.read_excel() — never rely on inference
• [ ] No raw pd.DataFrame crossing module boundaries without documented column contract
• [ ] Merge/join results checked for unexpected nulls and row count changes
• [ ] model_config = {"strict": True} on all Pydantic boundary models
• [ ] No inplace=True — deprecated, returns None, causes silent bugs
• [ ] Notebook logic that survived 3+ uses extracted into tested modules

Gotchas

| Trap | What to do instead | |---|---| | df["a"]["b"] = x (chained indexing) | df.loc[:, "b"] = x — chained indexing silently fails | | .apply(lambda) on large frames | Vectorized ops first; .apply() only when no vectorized path exists | | pd.merge() without post-check | Assert no unexpected nulls or duplicate keys after merge | | df.drop(..., inplace=True) | df = df.drop(...) — inplace is deprecated and returns None | | Bare pd.read_csv(path) | Always pass dtype= to prevent silent type inference errors |

Decision Table

| Task | Use | Not | |---|---|---| | Tabular < 1M rows | pandas | Polars (overhead not justified) | | Tabular > 1M rows or need speed | Polars | pandas | | SQL-like analytics on local files | DuckDB | Loading everything into pandas | | Read-only TOML config | tomllib (stdlib, binary mode "rb") | tomlkit | | Read/write TOML preserving comments | tomlkit (text mode) | tomllib |

Module Layout

etl/
├── ingest.py      # raw data loading (boundary)
├── validate.py    # schema validation (boundary)
├── transform.py   # business logic (typed core)
├── load.py        # output writing (boundary)
└── types.py       # shared typed models