Prompting
Change the mental model
Section titled “Change the mental model”When you open a browser, you think in steps. Go to this site, click this, scroll there, copy that. You’ve been doing it so long it’s automatic — a mental filter that converts what you want into how to get it.
With UseBrowser, drop the filter. You’re not operating a browser anymore. You’re describing what you want from the internet.
The question to start with: “What do I want from the internet today?”
Not “go to Amazon and search for…” — that’s the old filter. Instead:
“I want the cheapest DDR5 RAM I can get from any Malaysian marketplace, shipped within a week.”
Claude figures out where to look, what to compare, how to filter. You described the intent. The agent handles the browsing.
Be specific about what, open-ended about how
Section titled “Be specific about what, open-ended about how”The best prompts are precise about the outcome but leave room for the agent to figure out the approach:
Both get you a price list. But the second one lets Claude optimize — it might find a faster extraction method, skip irrelevant results, or discover a better sort than you’d have used.
Be specific about:
- What you actually want (intent)
- What the result should look like (format, fields, structure)
- Any constraints (price range, region, date, count)
Leave open:
- Which sites to visit (unless you care)
- How to navigate or extract
- What order to do things in
Examples that work
Section titled “Examples that work”Research:
“Compare the top 5 project management tools for small teams. Pricing, key features, free tier limits. Save a comparison report.”
Shopping:
“I want to buy a mechanical keyboard under RM 300. Find options across Shopee and Lazada with Cherry MX switches. Rank by value.”
Data extraction:
“Get all restaurants within 2km of KLCC on Google Maps with ratings, phone numbers, and opening hours.”
Account tasks:
“Check my GitHub notifications. Mark all as read. Open any PR reviews assigned to me in new tabs.”
Content creation:
“Read this article, summarize the key points, and draft a LinkedIn post about it.”
Chain workflows creatively
Section titled “Chain workflows creatively”Don’t stop at single tasks. Chain them:
“Research the top 5 VPN services. Then go to each one’s pricing page and extract the monthly cost for their cheapest plan. Save everything in one comparison table.”
“Find the cheapest flight from KL to Tokyo next month on Google Flights. Then check the same dates on Skyscanner. Tell me which is cheaper.”
“Go through my Gmail and find all receipts from this month. Extract the amounts and vendors. Save as a CSV.”
The agent can navigate multiple sites, cross-reference data, and compile results — all from one prompt. The more you describe what you want the final output to look like, the better the result.
From prompts to skills
Section titled “From prompts to skills”When a prompt works well, capture it so you never type it again.
/ub:learn — Turn what just worked into a skill
Section titled “/ub:learn — Turn what just worked into a skill”After Claude completes a task successfully:
- Type
/ub:learnin the terminal - Claude reads the conversation, extracts what worked, and generates a reusable Playwright script
- Next time, just describe the task and Claude uses the skill automatically
Push it further: After the first skill is generated, ask Claude to optimize it. “Can you make this faster? Can it handle pagination? What about error cases?” Two or three rounds of refinement turns a rough skill into a production-quality automation.
/ub:record — Show Claude, then let it learn
Section titled “/ub:record — Show Claude, then let it learn”For tasks that are easier to show than describe:
- Type
/ub:recordin terminal - Do the task in the browser — Claude captures everything (network requests, DOM changes, page state)
- Say “done”
- Claude asks clarifying questions and generates a reusable script
See Skills for the full details on both commands.