LLMent Picker nga Henning Pähtz

LLMent Picker (say: el-el-ment) sees nothing until you click, sends nothing anywhere and runs only on the page you point it at. One click copies exactly what an AI agent needs to know which element you mean – and, unlike an annotation workspace, it stays out of your way: no panel, no comments, no all-sites permission.

Working on a web page with an AI assistant, you hit this constantly: "I mean the second box under the heading" is clear to a person and useless to a tool. LLMent Picker turns one click into an unambiguous reference in your clipboard:

1. the full page URL
2. a CSS selector that matches exactly this element, e.g. #bereiche .values > .value:nth-of-type(2) > p
3. if you selected text first: that text in quotes – "this passage is meant"
4. with Alt+click: a screenshot of the element and its surroundings, plus viewport and pixel ratio – for "this overlaps", "this is misaligned". By default the PNG goes to your downloads folder and its path into the clipboard text – token-friendly, readable by any chat, croppable at full size (setting: file, clipboard with image and text in one Ctrl+V, or both). Elements larger than the window (long tables, wide tables in scroll containers) are captured whole, stitched from several tiles. Or hold the left button and drag a region: the lasso captures exactly that area, the selector names the container that covers it best.
5. with Ctrl+click: the element's HTML as currently rendered, saved as a file in your downloads folder, with the CSS context you would see in the inspector (matching rules with file and media query, effective layout values). The path goes into the clipboard – Claude Code reads it straight from disk. For content that only exists logged in or after rendering: timetables, tables, portals.

Selectors are readable for humans and unambiguous for tools: IDs and meaningful class names serve as anchors, positional indices are used only where needed, state and animation classes are ignored, and every selector is verified against the page before copying (exactly one match).

Usage
• Select text, then icon or shortcut → element containing the selection + text copied immediately
• Nothing selected: icon or shortcut → picker starts; the element under the cursor gets an outline; click copies
• Alt+click → with screenshot · Ctrl+click → as HTML file · both → both · drag → lasso region as screenshot · R → record the interaction for 5 s (hover rules, transitions, state diff, DOM changes, frame timing, contact sheet)
• Hold Shift → multi-selection: Shift+↑/↓ or Shift+click extend to neighbouring sections, any other element joins individually, a container becomes the context of the inner elements; Enter or a click copies everything with one screenshot
• Tap Alt, Ctrl or R to arm a mode for the next click – no need to hold the key while aiming
• Every copy starts with a label such as [LLMent: element, shot:file] and a context line (browser, viewport, pixel ratio, dark/reduced-motion) – an agent recognises the format instead of guessing; every part can be switched off
• Right-click on the page → "Copy this element" / "Copy with screenshot" / "Save as HTML file" / "Record interaction here", without picker mode
• Right-click the icon → start the picker with a preset
• Works inside iframes; the clipboard says which frame

Privacy: only the activeTab permission – the extension does nothing until you invoke it and has no access to other pages. Screenshots and HTML files are created only on your click and stay on your device (downloads folder or clipboard). No network access, no data collection. Source on GitHub, MIT licence.