browser-screenshot

A pipeline for capturing focused, region-specific screenshots from web pages rather than noisy full-page dumps. The workflow moves through four stages: research the best page to capture, navigate to it in Chrome, locate the target region via a CSS selector, and capture a cropped screenshot of just that region. A hard rule forbids emitting an uncropped full-viewport or full-page screenshot as a final result, since navigation bars, sidebars, and ads make such images unsuitable as article illustrations. The prerequisite is the agent-browser tool with Chrome running under remote debugging.

A research-first principle drives the early steps: the browser is treated as a capture tool, not a browsing tool. Text-based tools (WebSearch and WebFetch) are used first to find candidate pages, read their content, and decide which are worth screenshotting, eliminating poor candidates cheaply before any browser navigation. The skill also defines when to skip straight to Chrome instead — when the target platform gates content behind a login wall (Reddit, LinkedIn, X/Twitter), when the user names a platform with a clear search need, or when WebFetch returns blocked or incomplete content such as a 403 or CAPTCHA. Guidance on what makes a good source emphasizes content over chrome: blog post sections with a heading and a paragraph or two, a single chart, a HuggingFace model-card header, or a key quote, while avoiding busy landing pages dominated by buttons and forms.

Operationally it always starts by listing existing browser tabs to reuse logged-in sessions, then navigates by input type — direct URL, search query, or a constructed platform search URL (patterns are given for Reddit, X/Twitter, LinkedIn, Hacker News, GitHub, and YouTube). After navigation it waits for the page to settle with networkidle, noting that some sites never reach that state and a fixed wait is a safe fallback. The critical region-location step centers on discovering a precise DOM selector that wraps exactly the content to capture, beginning with an annotated screenshot to understand page layout.