pdf-extractlisted

# doc-extract — document → markdown (MinerU, local GPU) Converts PDFs and Office/scanned documents to clean markdown + extracted tables using MinerU on this machine (torch-ROCm, AMD GPU). Output is **text**, so it's cheap in the model context — this is the default document reader. Reach for pixel-read only when layout itself carries the meaning. Tool entrypoint: ```bash bash "$CLAUDE_PLUGIN_ROOT/bin/nw" doc <file.pdf> # → ./mineru-out/<name>/... markdown bash "$CLAUDE_PLUGIN_ROOT/bin/nw" doc <file.pdf> <outdir> # custom output dir ``` After it runs, read the produced `*.md` under the output dir. MinerU also writes extracted tables and a content-list JSON alongside the markdown. ## Why text-first (the token argument) MinerU emits markdown; pixel-read holds pixels. A screenshot in context costs many× the tokens of the same content as markdown. So for the typical document — reports, contracts, statements, papers: mostly prose + tables — MinerU is both cheaper and more faithful. Default here; escalate to pixel-read only for the layout-bound minority. ## When to use vs pixel-read | Document is… | Reader | |---|---| | prose, tables, statements, papers, contracts | **doc-extract** (text, cheap) | | charts/infographics where the figure is the point | **pixel-read** | | scanned form whose spatial layout = data | **pixel-read** | | scanned text pages (no layout meaning) | **doc-extract** (MinerU OCRs) | ## Rules - **Text-first.** Don't pixel-read a document tha