Browser Research Skill

Research web pages using agent-browser CLI and summarize content. A supplied URL alone does not require this skill when ordinary fetching already answers the question. Explicit user instructions take precedence over workflow preferences.

Usage

/browser-research <URL> [research topic or question]

Workflow

1. Read the page (default — no browser)

agent-browser read "<URL>" --max-output 20000

read never launches Chrome. It negotiates Accept: text/markdown, tries the .md variant, looks for the nearest llms.txt, and falls back to readable text extracted from the HTML. The markdown it returns keeps headings, bullet lists, and code blocks, so it is both cheaper and more faithful than scraping rendered text.

Useful variants:

Purpose	Command
Read one section only	agent-browser read "<URL>" --filter "<heading text>"
See the page structure first	agent-browser read "<URL>" --outline
List a docs site's llms.txt links	agent-browser read "<URL>" --llms index

Go to step 2 when read exits non-zero (e.g. Read failed with HTTP 403 — some sites block non-browser clients), returns nothing useful, or the content only appears after JavaScript runs. Interaction (clicking, dismissing banners, following links, tabs) always requires step 2.

2. Fall back to a browser session

Use a session owned by this task; do not close an unrelated existing session. Consult agent-browser --help for session selection when isolation is needed. Open the page:

agent-browser open "<URL>" && agent-browser wait --load networkidle --timeout 15000

If open fails, verify the URL and retry once when useful. If it still fails, report the inaccessible source and continue with other relevant sources when they can answer the question.

If wait times out: proceed anyway — the page may still be usable.

Next, choose the extraction method based on your purpose:

Purpose	Command	When to use
Read the rendered page as markdown	agent-browser read	No URL argument: reads the active tab's rendered DOM, including client-side updates and auth state
Read article/docs text	agent-browser eval "document.body.innerText"	Plain text of the rendered page. Includes navigation chrome and loses structure
Understand page structure	agent-browser snapshot -c	Need to see layout, navigation, or element refs for interaction
Find interactive elements	agent-browser snapshot -i -c	Need to click links, buttons, or fill forms
List links with URLs	agent-browser snapshot -i -c -u	Need link destinations (href) without an extra get attribute round-trip

Use snapshot only when you need structure or element refs.

For large pages, append --max-output 10000 to prevent token explosion:

agent-browser eval "document.body.innerText" --max-output 10000

If a cookie consent banner or overlay blocks content, dismiss it first:

agent-browser snapshot -i -c   # find the accept/close button ref
agent-browser click "@ref"     # dismiss the banner

Then proceed with the chosen extraction method.

3. Close when done

Only needed when step 2 opened a session:

agent-browser close

Deeper investigation

Most research finishes with the steps above. Read reference.md when you need any of these:

• The full read option reference (--raw, --require-md, --llms full, --timeout)
• Text from a specific element, page metadata, or finding elements by role/text/label
• Scrolling to load more content on long pages
• Watching network traffic to locate the JSON endpoint behind a SPA
• Following links, going back, or researching several URLs in tabs
• Saving the page as a PDF
• The full snapshot flag reference and other commands

Critical Rules

• Prefer read for text retrieval. Open a browser directly when interaction or JavaScript rendering is already known to be necessary.
• Always close a session you opened — every open must have a matching close.
• Read-only by default — never submit forms or enter data. Clicking is allowed only for passive navigation: dismissing cookie/consent banners, following links, expanding collapsed sections, or switching tabs. Do not click buttons that trigger writes, purchases, or state changes.
• No guessing — do not fabricate or assume page content; only report what read/snapshot/get/eval return.
• Authentication pages — do not attempt to authenticate as part of research. Report the inaccessible source and use relevant public sources if sufficient; ask for accessible source material when the answer depends on the protected page.
• Minimize tokens — prefer read (or read --filter) over full-page extraction, and use --max-output for large pages.

Output

Lead with the answer to the research question, then the supporting points and the source URLs the reader may want to follow. Respond directly in chat unless the user requested a saved artifact.