How to Build Your Own AI Agent Safely in Claude
Use Claude in Chrome as a safe AI agent for booking, shopping, form-filling, and research. The safe setup, starter prompts, and why this beats open-source alternatives.
You can give Claude access to your browser and let it book restaurants, compare prices, fill out forms, and run weekly research. It works right now, no coding required. But most guides skip the part where the AI accidentally buys something or submits the wrong form. This one does not. Here is how to set it up safely, with the exact prompts to copy.
- A safe Claude-in-Chrome setup that asks before every action
- 4 copy-paste prompts: booking, price comparison, form filling, weekly research
- Why Claude beats other browser agents on safety
- How to reframe this for business task automation
The safe setup: 3 things before you start
Claude can control your browser through the Claude desktop app and Chrome extension. That means it can click buttons, fill fields, and navigate pages on your behalf. Powerful, but also risky if you skip the safety layer.
- Get a paid Claude plan. Computer use is available on Pro ($20/mo) and Team plans. The free tier does not include it.
- Install the Claude Chrome extension. This is how Claude sees and interacts with your browser tabs. Install it from the Chrome Web Store and connect it to your Claude account.
- Turn on approval-before-action mode. This is the critical safety step. In Claude settings, enable the option that requires your explicit approval before Claude clicks, types, or submits anything. Without this, Claude will execute actions immediately. With it, Claude shows you exactly what it plans to do and waits for your OK.
The approval step is non-negotiable. An AI agent without a confirmation gate is a liability, not a tool.
Prompt 1: Restaurant booking (the mega prompt)
This is the full-featured prompt. It handles the entire OpenTable booking flow: searching, filtering by your preferences, selecting a time slot, and filling the reservation form. Every step pauses for your approval.
You are my booking assistant. Use Chrome to complete this task.
TASK: Book a dinner reservation on OpenTable.
DETAILS:
- Restaurant: [RESTAURANT NAME] in [CITY]
- Date: [DATE]
- Time: around [TIME], flexible 30 minutes either way
- Party size: [NUMBER]
- Seating preference: [indoor/outdoor/bar/no preference]
RULES:
1. Navigate to OpenTable and search for the restaurant.
2. If the exact time is unavailable, show me the closest available slots and WAIT for my choice.
3. Before clicking "Complete Reservation," show me the full reservation summary (restaurant, date, time, party size, any notes) and WAIT for my confirmation.
4. Do NOT submit any form without my explicit approval.
5. If the restaurant is not found, suggest 3 alternatives in the same area with similar cuisine and WAIT.
6. Fill in my contact info only after I confirm the details:
- Name: [YOUR NAME]
- Email: [YOUR EMAIL]
- Phone: [YOUR PHONE]
After booking, confirm the reservation number and any confirmation email.Replace the bracketed fields with your actual details. The key design pattern: every decision point includes "WAIT for my confirmation." This is how you keep the agent useful without giving it unsupervised authority.
Prompt 2: Price comparison across sites
Use this when you need to compare a product across multiple retailers. Claude opens each site, extracts the price, and builds a comparison for you.
You are my shopping research assistant. Use Chrome to complete this task.
TASK: Compare prices for [PRODUCT NAME / MODEL NUMBER] across these sites:
1. Amazon
2. Best Buy
3. Walmart
4. [ADD MORE SITES]
FOR EACH SITE:
- Search for the exact product
- Record: price, shipping cost, estimated delivery date, seller rating if available
- Screenshot the product page for my reference
RULES:
1. Do NOT add anything to a cart or start any checkout process.
2. If the exact model is not found on a site, note "not available" and move on.
3. After checking all sites, present a comparison table sorted by total cost (price + shipping).
4. Flag any suspicious listings (third-party sellers with low ratings, refurbished sold as new).
5. WAIT for my decision before taking any further action.Prompt 3: Form filling
For repetitive applications, registrations, or intake forms. Claude fills the fields and waits for you to review before submitting.
You are my form-filling assistant. Use Chrome to complete this task.
TASK: Fill out the form at [URL].
MY INFORMATION:
- Full name: [NAME]
- Email: [EMAIL]
- Phone: [PHONE]
- Address: [ADDRESS]
- [ADD ANY OTHER FIELDS SPECIFIC TO THIS FORM]
RULES:
1. Navigate to the URL and identify all form fields.
2. Fill each field with the matching information above.
3. For any field you are unsure about, leave it blank and ASK me what to enter.
4. After filling all fields, show me a summary of what you entered and WAIT for my approval.
5. Do NOT click Submit, Send, or any equivalent button until I explicitly say "submit."
6. If the form has multiple pages, pause at the end of each page for my review.Prompt 4: Weekly research roundup
Set this up as a recurring task. Claude visits your specified sources, pulls key updates, and compiles a brief for you.
You are my research assistant. Use Chrome to complete this task.
TASK: Weekly research roundup for [YOUR TOPIC / INDUSTRY].
SOURCES TO CHECK:
1. [SITE/URL 1 - e.g., TechCrunch, specific subreddit, industry blog]
2. [SITE/URL 2]
3. [SITE/URL 3]
4. [SITE/URL 4]
5. [SITE/URL 5]
FOR EACH SOURCE:
- Check the last 7 days of content
- Extract the 3 most relevant articles or discussions for my work
- For each: title, one-sentence summary, link, and why it matters to [YOUR CONTEXT]
OUTPUT FORMAT:
- Group by source
- End with a "Top 3 This Week" section picking the most actionable items across all sources
- Flag anything that requires immediate attention or action
RULES:
1. Only report content from the last 7 days.
2. Do NOT sign up for anything, create accounts, or bypass paywalls.
3. If a source is unreachable, note it and move on.Why Claude over other browser agents
There are other tools that let AI control your browser. Open-source options like OpenAdapt and browser-use exist. Some work well for developers comfortable debugging Python scripts. But for most business users, Claude has three advantages.
- Built-in safety guardrails. Claude is trained to refuse harmful actions, ask for clarification, and default to caution. Other agents execute whatever you tell them with no friction.
- No code required. You write plain English prompts. Open-source browser agents require Python setup, dependency management, and debugging. Claude works from a text box.
- Approval mode is native. The pause-before-action flow is a first-class feature, not a hack. You see what Claude plans to do, approve or edit, then it proceeds. This makes it usable for real tasks involving money and personal data.
The tradeoff is cost. Claude Pro is $20/month. Open-source agents are free. If you are a developer who wants full control and zero cost, the open-source path works. If you are a business owner who wants to automate tasks this week without writing code, Claude is the faster path.
For business owners: what this actually automates
Stop thinking about AI agents as tech demos. Think about the repetitive browser tasks you or your team do every week.
- Client intake forms. If you fill out the same vendor portals, client onboarding forms, or compliance documents repeatedly, the form-fill prompt handles it.
- Competitive price monitoring. The price comparison prompt works for checking competitor pricing, supplier costs, or marketplace listings on a schedule.
- Industry monitoring. The weekly research prompt replaces the 2 hours you spend scanning industry news, competitor blogs, and community discussions.
- Booking and scheduling. Restaurant reservations are the example, but the same pattern works for booking meeting rooms, scheduling demos, or reserving event spaces.
The ROI question is simple: if you spend more than 30 minutes a week on a browser-based task that follows a repeatable pattern, an AI agent can probably do it in 5 minutes with your approval on each step.
- Want AI automation built for your business workflows? See what jynlab builds
- Not ready for custom builds? Copy the prompts above and start with one task this week.
Builders are already on the list. New guides and teardowns, delivered when they ship.