I Built an AI Price Tracker That Caught a Fake "Cheapest" Deal

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: I Built an AI Price Tracker That Caught a Fake “Cheapest” Deal · Channel: Prompt Engineer 48 · 07:52 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, Task and workflow automation

Summary

The creator compares two runs of Claude Code researching Logitech MX Master 3S prices and reviews: one without a scraping skill and one using Decodo’s web scraping API. The skill-enabled agent retrieves data from Amazon, Walmart, Target, Reddit, Google, and YouTube, while also distinguishing new offers from a cheaper used listing and identifying a missing USB receiver in the recommended deal. The creator then turns the discovered API templates into a Streamlit price-tracking app that compares store offers, summarizes price differences, and triggers an alert at a chosen target price. The video also notes the skill run took longer and cost more in model tokens, but says the free plan’s request allowance was sufficient for the demo.

Key points

The research brief asks for prices and sellers across three stores, recent Amazon reviews, Reddit discussions, Google results, YouTube reviews, a verdict, and a dashboard.
Without scraping support, the agent encounters missing Amazon prices, a Walmart CAPTCHA, and blocked Reddit access, and produces a partial report.
A manual Amazon template request reveals a $73 used offer, illustrating why the app must compare new-condition offers separately.
With the Decodo skill loaded, the agent uses 14 templates and about 70 requests to collect information across the requested sources.
The report identifies Amazon’s $89.99 offer as sold by Amazon, while the Walmart and Target listings are from marketplace sellers.
The recommended Amazon listing is Bluetooth-only and lacks a USB receiver, an important caveat for the buying decision.
The skill-enabled run took about 25 minutes and used about $3.36 in model tokens, compared with three minutes and $0.76 without the skill.
The creator builds a Python and Streamlit app that normalizes store results into shared fields and checks prices across Amazon, Walmart, and Target.
The app compares the lowest price against a user-set target and saves each check; the demo shows an alert when the target is raised to $88.
The creator says the free plan includes up to 2,000 requests and demonstrates a separate YouTube template returning recent channel videos.

Chapters

Start	Chapter	Summary
00:00	The research brief and the scraping challenge	The creator introduces a two-run Claude Code comparison and explains why the requested shopping research is difficult to gather from websites directly.
01:06	Testing the API and configuring the agent	A manual Amazon request reveals a misleading used offer, then the creator explains how to install the agent skill and provide its token.
02:06	Comparing the two research runs	The agent without the skill returns partial results, while the skill-enabled agent uses templates to build a fuller report and buying verdict.
04:40	Building the price-tracking app	The creator weighs the longer, more expensive research run against its results and builds a Streamlit app from the API templates.
06:19	Price alerts, usage, and setup steps	The app demonstrates its price summary and target alert, followed by the creator’s usage figures and instructions for trying the workflow.

Quotes

The cheapest Amazon offer is $73, but read the line next to it, used, very good, from a returns seller. — 01:36

That’s the difference between a guess and a decision. — 04:10

The skill run took about 25 minutes. The other one, three. — 04:40

No fake price drop, just a rule running on the real price. — 06:50

Concepts and entities

Claude Code, Decodo, Decodo web scraping API, Decodo agent skill, Logitech MX Master 3S, AI Price Watch, Python, Streamlit

Full transcript (original language)

00:00 I gave Claude Code one big job. Twice, same prompt, same model, same tools. The only difference, on the right it has Decodose web scraping API, on the left it doesn’t look at what came back. Let me walk you through how we got here. Here’s the job, a buyer’s research brief for one product. The Logitech MX Master 3S five parts, the new price at Amazon, Walmart, and Target with the seller, the 10 latest Amazon reviews, what Reddit is saying, the top 10 Google results, and

00:31 the most viewed YouTube reviews, then a verdict, buy now or wait, and a dashboard. That’s a real project, it’s what every shop owner, reseller, and deal hunter does by hand. And every one of those five sources fights back. Amazon shows bots a page with no price. Walmart throws a CAPTCHA. Reddit refuses the request. Google and YouTube build a page with JavaScript. A smart model can’t prompt its way past any of that. So, here’s the setup and it’s two steps. Step one, in the Decodose dashboard, the free web scraping API plan, it’s active

01:06 on my account up to 2,000 requests for a year. No credit card. In the playground, your authentication token is right there, one click to copy. But before you hand anything to an agent, try one request by hand. It takes about a minute. Instead of writing a scraper, I pick a template, Amazon pricing. All it needs is the product’s ASIN, and I’ll ask for the result as markdown. Send request. The task is created and a few seconds later, response code 200

01:36 with every offer for this mouse. And this is where it got interesting. The cheapest Amazon offer is $73, but read the line next to it, used, very good, from a returns seller. If my app just grabbed the lowest number, it would have told you Amazon wins, and that would be wrong. So, anything I build has to compare new offers only. Step two, give the token to your agent. Decodo publishes agent skills for Claude Code, the coding agent, at the marketplace.

02:06 Install the plugin and put the token in an environment variable. That’s it. The skill teaches the agent when to reach for the API and which template to use. Now, the test. Two fresh folders, same prompt, word for word, and it never says Decodo. Same model, same tools, both have web search and web fetch. The only difference is whether the skill is loaded. And the go on the left, the agent without the skill tries everything. Amazon, no price on the

02:37 page. Walmart, captcha. Reddit, blocked. It tries a price history site, deal articles, a different search tool. After 2 minutes 56, it gives up and writes what it has. On the right, the agent checks its token and goes straight to templates. Amazon search, Amazon product, Walmart search, Target, Google, Reddit, subreddit, Reddit post, YouTube search, YouTube metadata. 14 different templates, about 70 requests. It hits errors, too, like a bad location

03:08 parameter, and it fixes them on its own. Here’s the brief without the skill. A warning on top, partial data. Amazon, unavailable. Walmart, captcha. Reviews, unavailable. Reddit, unavailable. The Google section isn’t Google. It’s a substitute and the verdict wait is built on deal articles that are 6 weeks old. To be fair, it was honest about all of it. It just had nothing to work with. And here’s the brief with the skill. Every new offer across all three stores,

03:39 cheapest first, each one marked first party or third party. Amazon itself sells it for $89.99. Walmart’s main listing is $104.99 from a marketplace seller. Target, $103.60 sold by SpaceBound, not Target. The agent on the left actually got that seller wrong. Keep scrolling. 10 real Amazon reviews with stars and dates, Reddit threads with real scores, and a sentiment read that found a pattern. Left click switches failing after 18 months to 2

04:10 and 1/2 years. The Google top 10 and the five most viewed YouTube reviews with view counts starting at over a million. So, compare the verdicts without the skill weight, low confidence with the skill buy now from Amazon 89.99 sold by Amazon itself with a 30-day return window and a catch I would have missed. That listing has no USB receiver. It’s Bluetooth only. That’s the difference between a guess and a decision. Now, the honest part. The skill run took about 25

04:40 minutes. The other one, three. It used more model tokens, too, about $3.36 against $0.76. It explores, it retries, it checks its work. The API side, about 70 requests, fits easily inside the free plan’s 2,000. So, when is it worth it? When the answer has to be right. Price tracking, competitor research, review mining, anything you’d run every day. And once you know which templates work, you don’t need the agent to explore again. You put those calls in a script and it runs in seconds. So, that’s exactly what I did next. I took the

05:12 templates the agent found and built a small app you can use every morning. AI Price Watch, Python and Streamlit on top, Decodable underneath. Here’s the code. One small request function and the token comes from an environment variable, never from the code. I built the three store adapters on top of it with Claude code using the request formats from Decodable’s agent skill docs. Each adapter uses a product template and one run of the script checks all three stores. Target takes the longest, so that waiting time is

05:46 sped up and there’s the answer. Walmart 8760. The trick is making the stores comparable. Amazon gives me a list of offers. Walmart gives me a product object. Target’s parsed data doesn’t even have a price, so that adapter reads it from the rendered page. They all end up as the same row, store, price, seller, condition, stock, link, and the time it was checked. Now the app check prices. Now Amazon and Walmart come back in about 5 seconds. Target needs a full browser render, so it lands last. Three

06:19 out of three, all live, all new condition. Then the AI summary. Python has already worked out the lowest price and the gap, so the model isn’t doing maths. It explains Walmart is cheapest by $2.34. Both of the cheapest offers are third-party marketplace sellers, and the target isn’t reached yet with source links. And the alert, my target is $85. The lowest listed price is $87.60, so nothing fires. Raise the target to 88

06:50 and it trips. Target price reached. No fake price drop, just a rule running on the real price. Every check is saved. The lines are flat because these stores generally kept the same price between my checks, and the app says exactly that. And here’s my dashboard from when I recorded the app, 89 successful requests and 10 cents of the free plans, $1 allowance. The app runs comfortably on the free plan. And it’s not only prices. There are over 100 templates, Google,

07:21 Reddit, TikTok, Airbnb, YouTube. I pointed the YouTube channel template at my own channel and got my latest 20 videos back as a spreadsheet with titles, views, and dates. Same token, same one request. Here’s how you do it yourself. One, grab the free scraping API plan and copy your token. Two, install the Decodable skill and set the token. Three, give your agent a real job like this one, then look at what it used and turn it into a script. The free

07:52 scraping API plan gives you up to 2,000 requests, no credit card, and the link is in the description. When you’re ready to scale, my code prompt, that’s p r o m p t, takes 10% off the web scraping API. See you in the next one.