Since I started using OpenClaw, I've been tinkering with various workflows — except when I'm too busy or tired from work. Recently, I've wanted to share some of my experiences from time to time.
This time, the project is: a shopping assistant that started as a simple shopping list and gradually grew into a tool that can search Xiaohongshu (Little Red Book), compare prices, and add items to the shopping cart.
More and More Things to Buy
Recently, there's been a lot to buy for the household — children's picture books, microscopes, roller skates, and various daily necessities. Every time I need to buy something, I have to switch between Taobao, JD.com, and Xiaohongshu to compare prices, check reviews, and pick the right model. It's quite tedious.
I had previously built an Agent called the "Shopping List Manager" to help me keep track of things I wanted to buy. Whenever I thought of something, I'd just tell it, and it would categorize and organize everything — basically a shopping memo.
But a memo is just a starting point. Once you have a basic function, you can't help but wonder — can it check prices for me? Can it look up reviews on Xiaohongshu? Can it add items directly to my shopping cart?
First Evolution: From Memo to Auto Price Comparison
After using the shopping memo for a while, I started feeling that just recording things wasn't enough. I knew what I wanted to buy, but I still had to look up prices myself.
Then one day, my Skill Researcher (an Agent that automatically searches for new Skills every day) found a price comparison Skill. After installing it, it could search prices across Taobao, JD.com, and Pinduoduo, telling me roughly what options were available and what price range to expect.
With price comparison, the shopping assistant evolved from "only recording, no searching" to "recording and searching."
Second Evolution: Bringing Xiaohongshu Into the Mix
I had price comparison, but I mainly relied on Xiaohongshu for reviews and recommendations. Just knowing the price range wasn't enough — I also wanted to know "which brand is good" and "what pitfalls others have encountered."
There was no ready-made Skill for Xiaohongshu, but I was already using OpenClaw's browser Skill — through Brave Browser's CDP protocol (essentially opening the browser's debugging port), the Agent can open a browser and operate web pages like a human. I had the Agent use this capability to search Xiaohongshu and read what users were saying.
Later, I used the Skill Creator to slowly refine a Xiaohongshu search Skill specifically for search and content extraction. It wasn't written in one go — it went through multiple rounds of debugging, trial and error, and revisions before it stabilized.
At first, the Agent actually fooled me once. It said "I checked for you" and gave detailed recommendations that looked quite convincing. But when I had it reopen Xiaohongshu and search for actual posts, I realized the previous recommendations were largely fabricated.
I also found issues with search sorting. Xiaohongshu defaults to sorting by likes, but highly liked posts aren't necessarily useful — many are fluff pieces or ads. The truly valuable information is often in the comments. Later, I required it to sort by comment count — posts with more comments have a higher probability of being genuine user sharing.
To prevent the Agent from making things up, I added a rule: recommendations must include evidence — which post said what, and what the post link is. From the very first time I had it search Xiaohongshu, I required it to include data sources. Over time, this rule became increasingly strict, evolving from simple screenshots to complete post links and citations.
Third Evolution: From Searching to Buying — Adding Items to Cart
Once when buying picture books, I really didn't feel like opening Taobao to search for each one and add them to the cart individually. I thought, can the Agent do this step for me?
At first, the Agent actually refused, saying it was unsafe and not recommended to directly operate shopping websites. I said adding to the cart is fine — I'll handle the payment. So I used the browser Skill again to create a Taobao shopping cart Skill. The Agent opens Taobao's website to search for products, compares prices and specifications in the search results, and adds suitable ones to the cart. The principle is the same as Xiaohongshu — all done through the browser Skill, just targeting a different website. After a few tries, it actually worked.
From memo to price comparison, to Xiaohongshu search, to Taobao cart — capabilities were connected one by one. First it could record, then search, then verify, and finally take action.
In Practice: Buying Picture Books
Buying picture books was the most complete workflow.
I didn't jump straight to having it search Taobao. Instead, I first used Gemini for a round of deep research — telling it my child's age and interests, and having it do a thorough research to recommend suitable picture books. After getting the recommendations, I passed the results to the shopping assistant to check editions, prices, and parent reviews.
The most crucial part was Xiaohongshu verification: sorting by comments, checking which recommendations were genuine, which looked like ads, and whether parents were complaining in the comments. Finally, it organized the recommendations into batches — what to buy first, what to buy later — and I had it add them directly to the Taobao cart. I checked the cart on my phone, the prices looked good, so I paid.
Individually, none of these steps are remarkable. But connected together, I no longer have to bounce between multiple apps.
Other Shopping Experiences
For the microscope, the Agent helped with research and recommendations, but I ended up buying a different model — its recommendations gave me a reference direction, and the final decision was based on my own judgment.
The roller skates were bought earlier than the picture books, before the cart-adding capability existed. The Agent helped search for brand recommendations on Xiaohongshu, and then I bought them myself on Taobao.
These three shopping experiences happened to fall into three stages: roller skates could only search Xiaohongshu, the microscope could search and recommend, and the picture books went through the full chain from research to order. Each evolution emerged naturally through actual use.
It started as just a shopping list, and gradually it learned to search, filter, verify, and add to cart. I originally just wanted to be a bit lazy, but the lazier I got, the more it could do — and in the end, it really did help me switch between fewer apps. I'll write a separate post about how to build it from scratch when I have time.
Top comments (0)