Drawing Analyzer

Construction drawings are the worst possible input for an AI. They are large PDFs full of symbols and cross references, where a single-dashed line (cold water) and a double-dashed line (hot water) look almost identical but mean completely different things. Uploading a 10 MB set directly fails, overcounts, and burns 60-70k tokens per question.

This skill fixes that with one principle: scripts do the cheap mechanical work, the model does the understanding. You run the indexing pipeline once per drawing set, which turns a huge PDF into small markdown indexes plus per-sheet text and images. Every later question reads those tiny indexes instead of the full PDF.

Do NOT write pattern-matching scripts to classify drawings (e.g. "if title contains 'electrical' then electrical"). Drawing formats vary too much. Classification, cross-referencing, and symbol reading MUST be done by you reading each sheet. Scripts only split, render, extract text, and extract vectors.

Setup (first time only)

pip install -r "$SKILL_DIR/scripts/requirements.txt"   # PyMuPDF

Mode A — INDEX (run once per drawing set)

Goal: turn merged.pdf into a analysis/ folder of small, queryable artifacts.