diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..59f8446
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,36 @@
+# CLAUDE.md
+
+## Project
+
+Accidental classifier — a small CNN that identifies sharp, flat, and natural symbols in cropped images of engraved music scores. Part of a larger OMR (optical music recognition) pipeline.
+
+## Key facts
+
+- **Model**: 3-block CNN (16→32→64 channels), ~115k params, 1×40×40 grayscale input, 3-class output
+- **Dataset**: Crops extracted from PrIMuS via Verovio rendering. ~5k per class after subsampling.
+- **Performance**: 100% on held-out test set (750 samples). The three classes are visually distinct enough that this is a solved problem.
+- **Seed**: 42. All splits use this seed for reproducibility.
+- **Split ratios**: 5% test, 10% val, rest train (stratified)
+- **Labels**: `["flat", "natural", "sharp"]` — alphabetical order, indices 0/1/2
+
+## Layout
+
+- `train.py` — training pipeline, defines `AccidentalCNN` and all data loading
+- `eval.py` — evaluates saved checkpoint on the test split (imports from train.py)
+- `extract_accidentals.py` — multi-font crop extraction from PrIMuS MEI files
+- `extract_fast.py` — faster single-font variant
+- `explore_dataset.py` — dataset statistics
+- `segment_test.py` — connected-component segmentation experiment
+- `model/accidental_classifier.pt` — saved weights
+- `crops/{flat,natural,sharp}/` — extracted crops (gitignored)
+- `dataset/` — raw PrIMuS data (gitignored)
+
+## Conventions
+
+- All Python scripts use `sys.stdout.reconfigure(encoding="utf-8")` for Windows compatibility
+- Hardcoded paths under `C:\src\accidentals\` — this is a local workstation project
+- No requirements.txt; deps are PyTorch, torchvision, NumPy, Pillow, OpenCV, and optionally Verovio + cairosvg for extraction
+
+## Remote
+
+- Gitea at `git.xml-pipeline.org/dullfig/accidentals`, branch `main`