pbi-context: Skill for Claude Code

Analyze a Power BI model

Target model: $ARGUMENTS

If no path was given, ask the user for the .pbit file, .pbip file, .SemanticModel/ folder, or project root folder before doing anything else.

Step 1 — Extract

Run the extractor once to generate the model's indexed output:

pbi-context -i "$ARGUMENTS" -o output

If pbi-context is not on PATH (not installed as a package), run it as a module from this skill's repo root instead — cli.py uses relative imports, so it must run with -m, not as a bare script:

python -m pbi_extractor.cli -i "$ARGUMENTS" -o output

run with working directory ${CLAUDE_PROJECT_DIR} (this skill's repo root), regardless of where $ARGUMENTS itself points on disk.

Add --lang es for Spanish documentation, or --index-format toon when the model is large and every token spent loading table/column/relationship listings matters.

The output lands in output/<model-name>/, where <model-name> is the input file's stem (for .pbit) or the .SemanticModel folder's name without the suffix (for .pbip).

Step 2 — Query, never load whole files speculatively

Use pbi-context --query output/<model-name> ... (or python -m pbi_extractor.cli --query ... with the same fallback rule as Step 1) instead of reading tables/<Name>.json or relationships.json directly — the resolver already handles TOON decoding for you, so you never need to think about {__toon, __fields, __rows} at all, regardless of which --index-format was used.

1. Always start with output/<model-name>/index.json (read the file directly, it's tiny). It has, per table: name, hidden/technical flags, column count, measure count, categories present. For "what's in this model" / "how many tables" questions, index.json alone is usually enough.
2. For a specific table's columns/measures/DAX, run pbi-context --query output/<model-name> --table "TableName" — returns one JSON object with a unified measures list (name, category, complexity, format_string, display_folder, formatted_expression). Add nothing else; don't read the raw tables/*.json file yourself.
3. For one specific measure once you already know the table and measure name, run pbi-context --query output/<model-name> --table "TableName" --measure "MeasureName" — cheaper than loading the whole table.
4. For "which tables have X measures/columns" / "find the measure/column that does Y" questions (something index.json alone can't answer — it would otherwise require opening every table file by hand), run pbi-context --query output/<model-name> --search-measures "keyword" or --search-columns "keyword" (optionally --category revenue etc.).
5. For relationship/join questions, run pbi-context --query output/<model-name> --relationships [--table "TableName"].
6. For a full human-readable narrative doc (the user wants something to paste elsewhere, not an answer to a specific question), read output/<model-name>/model_documentation.md directly.
7. metadata.json is the complete, unfiltered dump (every table, every column, every DAX expression). Only read it directly if the question genuinely spans the whole model and index.json + a couple of --query calls aren't enough — it's the heaviest file in the output.