Skip to main content
This scenario is what most people imagine when they hear “agent.” The model is given a goal that requires using the web like a person would: open a real page, read it, decide what’s relevant, extract specific data, and produce a structured artifact. The playwright-cli skill makes that one prompt away.

​
The prompt

Use the playwright-cli skill to open Hacker News (https://news.ycombinator.com),
find the top 5 stories whose titles mention AI, LLMs, agents, or Claude
(case-insensitive). For each: title, points, comment count, link.

Return the result as a clean Markdown digest with a single-line summary at the top.
Take a screenshot of the HN front page and include it in the output.

​
What happened

The playwright-cli skill gives the model high-level wrappers for the same Playwright primitives a human developer would use: navigate, wait for content, query selectors, extract text, screenshot. Combined with the live browser, the user can watch the agent operate the page in real time. On this run, the model:
  1. Opened https://news.ycombinator.com in the sandbox browser.
  2. Took a full-page screenshot of the front page.
  3. Extracted the visible post list — title, score, comment count, link, and rank — across the front pages it scanned (the summary mentions “trending across pages 1-3”).
  4. Filtered the list against the keyword set (AI, LLMs?, agents?, Claude, case-insensitive).
  5. Ranked the matches by points and picked the top five.
  6. Wrote a Markdown table with title / points / comments / link columns and a one-line summary at the top.

​
The output

HN AI digest result — quick summary on the left with emoji bullets, Markdown digest on the right The right-hand panel renders the Markdown digest:

​
Hacker News AI/LLM/Agent Top 5 Digest 🤖

Summary: Top AI stories today feature massive DeepSeek updates, Claude criticism, agent collaboration tools, and LLM-powered browser automation tools trending across pages 1-3.
#TitlePointsComments
1DeepSeek v4 …2,0291,546
…………
And the left side gives a quick scan with hand-picked emojis:
  • 🐳 DeepSeek v4 launching with 2,029 points (huge community buzz!)
  • 😤 I cancelled Claude criticism piece with 933 points
  • 🧠 Multiple agent tools trending (Karpathy’s LLM wiki, agent code testing)
  • 😔 Public sentiment piece about the AI industry’s reputation
The screenshot of the HN front page is also embedded in the message — visible at the top of the panel — so the user can verify the data was pulled from the real page, not hallucinated.

​
Why it works

This is the cleanest example in the course of the model using a tool, not just describing one:
  1. The playwright-cli skill gives it navigate / screenshot / extract primitives that map directly to the task.
  2. The live browser means the user can step in if something goes wrong — for example, if HN is down and the page doesn’t load, the user sees the same blank page the agent does.
  3. The terminal lets the model run filtering and ranking in Python after extraction, instead of trying to do it all from the LLM’s working memory.
The same prompt would work tomorrow against any other content site: Reddit, ProductHunt, Twitter/X (with login), arXiv, GitHub trending. The skill is the abstraction; the goal is interchangeable. Next: Building the invoice-builder skill → — the moment of leverage.