I used Claude Code to build Sitewatch — a monitoring platform that catches when websites return 200 OK but are actually broken. Not a side project. A real SaaS, in production, with paying users.
AI wrote roughly 60-70% of the codebase. Here are the things I wish someone had told me before I started.
1. Planning matters more than prompting
This is the single biggest lesson and the one I almost learned too late.
Early on, I'd open Claude Code and start prompting. "Build the check scheduling system." "Add retry logic." "Create the alerting pipeline." And it would. Working code, every time. The problem showed up three weeks in — I had a codebase full of features that each worked individually but didn't fit together. Different error handling patterns. Inconsistent data models. State flowing in three different directions.
I had to stop and rewrite the architecture. That cost me more time than if I'd planned it from the start.
Now my workflow is: plan first, prompt second. I think through data models, module boundaries, API surfaces, and error conventions before Claude Code writes anything. I write it down as a spec. Then I hand it pieces of the spec, one at a time.
The difference is enormous. Same tool, completely different outcomes.
2. The first 80% takes minutes. The last 20% takes your brain.
This is the productivity pattern nobody warns you about.
I'd describe a feature — "add multi-region check support with independent retry confirmation per region" — and Claude Code would produce a working implementation in minutes. Maybe 80% done. The remaining 20% — edge cases, failure modes, integration with existing code, the weird thing that happens when two regions disagree — that still required my experience.
That 80/20 split is incredibly valuable. But you have to know it's coming. If you assume the first output is done, you'll ship bugs.
3. Establish your patterns early — then Claude Code follows them
Once I had a consistent way of structuring API handlers, formatting errors, writing composables, and handling state — Claude Code replicated those patterns reliably. It picks up conventions from your codebase and follows them.
But it has to see them first. If your first few files are inconsistent, Claude Code will be too. Build the first few things yourself with intentional patterns. Let the AI learn from your codebase. Then it scales those patterns across everything else.
This is the secret to a Claude Code codebase that feels coherent instead of stitched together.
4. Domain expertise can't be delegated
Sitewatch needs to understand HTTP at a level most developers never think about. How CDNs rewrite MIME types. Why a 200 OK can mean "your site is completely broken." When a redirect chain is looping only for specific cookie states. Why the same page can return different content from different edge locations.
Claude Code could write the plumbing around all of this. But the actual detection rules — the logic that makes the product valuable — came from my head. That's not a limitation of the tool. That's the nature of domain expertise. AI handles the how. You provide the what and why.
5. Test generation is the most underrated feature
Everyone talks about Claude Code writing features. Nobody talks about how good it is at writing tests.
I'd point it at existing code and say "write comprehensive tests for this module, including edge cases." And it would generate tests I hadn't thought of. Race conditions I'd missed. Boundary values I'd overlooked. Error paths I'd assumed were unreachable.
For a solo builder who knows they should test more but never has time — this was the single biggest quality improvement.
6. Review intensity should scale with stakes
I don't hold Claude Code back from anything. It handles more than you'd expect, including complex logic. But my review process isn't uniform.
Boilerplate scaffolding? Quick scan. Database migrations? Check the schema. API route for a new feature? Read it carefully. Core detection logic that determines whether a site is flagged as broken? Line by line, every time.
The tool is good enough for everything. Your review discipline needs to match the risk.
7. Shipping faster means you need to verify faster
This one surprised me.
When you build with Claude Code, your development speed increases dramatically. Features that would have taken a week take days. But your production failure rate doesn't change — you're still deploying to the same CDNs, the same edge networks, the same infrastructure that's always been unpredictable.
So you ship more code, more often, to the same fragile production environment. Without better production verification, you're just creating more opportunities for things to break silently.
This is actually what motivated me to build Sitewatch in the first place. The faster I shipped, the more I needed to know — immediately — when something broke. Not "the server returned 200." Whether the site actually works.
What I built with it
Since the whole article is about building Sitewatch with Claude Code — here's what it actually does.
Sitewatch detects when your website returns 200 OK but is functionally broken. The kind of failures that traditional uptime monitors completely miss:
- Your deploy goes through but your CDN serves stale HTML referencing deleted JavaScript bundles — blank page, green dashboard
- Your JS file gets served with
Content-Type: text/htmland the browser silently blocks it — app looks loaded, nothing works - A redirect loop kills the page but monitoring only sees the first
301 - Your CDN returns a branded error page with
200 OKinstead of your actual content
What Sitewatch does:
- Verifies that assets your HTML references load with correct status and MIME type
- Follows redirect chains to completion
- Fingerprints content to detect silent changes
- Checks from multiple regions to catch CDN edge divergence
- Classifies root cause with stack-specific fix guidance
For developers: catches the deploy failures your uptime tool misses. Free tier, one site, 30 seconds to set up.
For agencies managing 25-100+ client sites: continuous verification across all sites, multiple regions, root cause classification. Stop finding out from client emails.
Pricing: Free (1 site) · $9/mo (25 sites) · $19/mo (100 sites)
That about wraps it up! If you found this useful, leave a comment or connect with me on Twitter | LinkedIn or visit my website.
Top comments (2)
This feels very real. I’ve had moments where AI helped me move fast… but later I realized I didn’t fully own what I shipped. That tradeoff is subtle but important.
So true... 🫡