An experienced web developer recounts using Claude Code (with Opus) and Pi (with GLM) to rebuild his side project GiftyWeddings.com, turning it from a simple gift registry into a full wedding website builder over 275 commits and about 10 days. He shares his evolving skepticism of AI coding tools, his iterative workflow (seeding the agent with hand-written code, reviewing output, testing in-browser, but skimming AI-generated tests lightly), and practical lessons: LLMs write overly verbose comments, giving Claude a headless browser to take its own screenshots dramatically improved its HTML/CSS output, sandboxing is essential since the agent ran destructive commands, and long sessions need context compaction. The tech stack stayed simple: Go backend, SQLite, htmx, vanilla JS, hosted on Fly.io.
Questions this post answers
How can giving Claude Code a headless browser improve its ability to write HTML and CSS?
Letting Claude start the server, add fake data through the app's own forms, and use a headless Chromium browser to capture and inspect PNG screenshots significantly improved its HTML and CSS output, compared to manually uploading screenshots. The tradeoff is higher token usage, so it's best limited to significant frontend changes rather than every commit. Developers refining AI coding workflows can find more practical agent tips on daily.dev.
What are the risks of running an AI coding agent like Claude Code without sandboxing?
An AI coding agent can run destructive commands with unset environment variables, such as executing rm -rf $SOMEVAR/*.png where SOMEVAR was empty, wiping out test files. This happened even inside a sandboxed container, underscoring the recommendation to always run coding agents inside a container or virtual machine rather than directly on a host system. Anyone adopting agentic coding tools can track safety practices like sandboxing on daily.dev.
Why do AI coding agents tend to write overly verbose code comments?
AI coding agents like Claude tend to default to writing excessively verbose comments unless explicitly told to be succinct, requiring repeated prompting to trim them down. In one case, a single cleanup commit reduced roughly 2500 comment lines down to 1500, showing how much unnecessary comment bloat can accumulate during iterative AI-assisted development. Developers tuning AI-generated code style can compare workflow tips on daily.dev.