visual-readlisted

# pixel-read — render to images, give Claude eyes Renders a URL, PDF, or HTML file to tiled JPEGs with pixelrag's `pixelshot`, then Claude reads the images directly. Use when the **layout itself carries information** that text extraction destroys. This is the costly reader — many× the tokens of markdown — so it is the exception, not the default. Tool entrypoint: ```bash bash "$CLAUDE_PLUGIN_ROOT/bin/nw" pixel <url|pdf|html> # → ./pixel-tiles/... JPEGs (paths printed) bash "$CLAUDE_PLUGIN_ROOT/bin/nw" pixel <url|pdf> <outdir> # custom output dir ``` It prints the produced tile paths. **Read those JPEGs** (Read tool) to see the page. ## Use ONLY when layout is the meaning | Input | pixel-read? | |---|---| | chart / infographic / data-viz where the figure is the point | **yes** | | scanned form whose spatial arrangement = the data | **yes** | | complex table that linearizes wrong in text | **yes** (that page) | | visual diff / "does this render correctly" | **yes** | | prose article, normal PDF, statement, paper | **no** → web-scrape / doc-extract | ## Rules - **Text-first, always.** Try web-scrape (page) or doc-extract (document) first. Escalate to pixel-read only when the text result is demonstrably missing layout-bound meaning — never as the opening move, never "to be safe". - **Scope the render.** Pixel cost scales with page count/size. Render only the page(s) that need eyes, not a whole document, when you can target them. - **Local, no index.**