Watch the video with chapters & transcript →
CAREERS
In short: A website change detector built in Claude Cowork, without writing code, monitors 69 job-listing URLs and has collected 2,002 jobs in its history.
Checking dozens of job boards daily is a chore that scales badly. Ani Björkström, a tech consultant based in Stockholm, replaced that routine by asking Claude Cowork to build a website change detector that scrapes 69 job-listing URLs, stores every posting in a searchable history, and emails a digest of new opportunities on a schedule.
You describe the app in plain language to Claude Cowork and answer the clarifying questions it asks before building. Claude responded to the prompt "create a website change detector" with key questions: how listings should be extracted, how email notifications should be delivered, how URLs get added, even which text editor is preferred.
Each question comes with a recommended option, so the whole specification is answered by multiple choice. While it builds, Claude shows its task list under a progress panel and writes real files to a folder on your computer, including a requirements file, environment example, gitignore, and Dockerfile.
Claude asks about extraction, delivery, hosting, and error handling before writing any code. These were the answers behind the finished job agent.
| Setup question | Option selected |
|---|---|
| How should listings be extracted? | Listing URL plus title, the recommended option |
| How should email notifications be delivered? | One digest per schedule |
| Which LLM API provider for extraction? | Anthropic Claude API, on an existing subscription |
| Where should the app run? | A Docker container on the user's own computer |
| Which extra behaviors are enabled? | Retry plus error alerting when scraping fails |
Three things: an Anthropic API key in the .env file, a Gmail address as the SMTP username, and an app password created in Google Cloud Platform as the SMTP password. If Docker is not installed, that is the fourth and final prerequisite before running the single Docker command that starts the container.
Once the container runs, the app is served from your own machine at a local HTTP URL. After several iterations with Claude, the result was a working website with a URL tab, a run button, and a history view.
Instead of visiting 69 URLs and reading every listing one by one, one click fetches all jobs and an email digest delivers the new ones. The history view held 2,002 jobs after three runs, and filtering by location reduced that to 384 Stockholm positions, each opening the original posting with its apply button one click away.
No. Claude Cowork asks multiple-choice questions, then generates every file itself, including the Dockerfile. Your work is answering questions and pasting credentials into the .env file.
In a Docker container on your own computer, which acts as the server. You access the app through a local HTTP URL once the container is started with a single Docker command.
By one digest email per schedule, so you are not pinged for every posting. If scraping a site fails, the app retries and sends an error alert, because retry plus error alerting was enabled during setup.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?