kg-edalisted
Install: claude install-skill TranBaDat2607/claude-code-kaggle-grandmaster
# /kg-eda
Run kgkit with the Bash tool, as `PYTHONPATH="${CLAUDE_PLUGIN_ROOT}" python -m kgkit <command>`: this command
pre-approves that form for the kgkit calls it makes. Submitting, pushing or publishing still asks (so does a
PowerShell `$env:` assignment).
Arguments: `$ARGUMENTS` (defaults: `data/train.*` and `data/test.*`; target/id from
`.kaggle-gm/competition.json`).
Run kgkit with the absolute "kgkit home" path from the session context (or `$KGKIT_HOME`):
1. Tabular: `python -m kgkit eda <train> --test <test> --target <y> --id-col <id> --out reports/eda.md`
then `python -m kgkit adv <train> <test> --drop <id>,<y> --save-p artifacts/adv_p_test.csv`.
2. Non-tabular (images/audio/text) or anything needing deeper investigation: delegate to the
`data-detective` agent with the dataset description.
3. Interpret every red flag for the user: what it means for validation, features, and leaks; the
concrete action for each.
4. The findings stay in `reports/eda.md`. In the workspace CLAUDE.md, add only what every session
must keep in mind (a leak, a shift, a trap in the data) to `## Current state`, one line each with a pointer
to the report: CLAUDE.md loads into every session, so it holds the conclusions, not the analysis.