Skip to content
Sweet Gum.
How we build
Our stack · AI

How We Use Claude and Claude Code

Where AI helps us build faster, and the gates that keep it from publishing anything nobody checked.

We use Claude, Anthropic's AI assistant, and Claude Code, its coding tool, to build websites, tools and content pipelines faster. We do not let it publish anything unchecked. Every output passes automated checks, and anything that matters passes a person.

That second sentence is the whole philosophy. AI is how we work, not what we sell. You are buying leads, calls and less busywork. Here is what the tooling looks like behind that.

Who this is for

You are a business owner who has heard "built with AI" and wonders what that means for quality. Or you are considering workflow automation and want to know how automated systems stay trustworthy. This page answers both.

How we use Claude Code to build

Claude Code works inside a project folder: it reads the code, edits files, runs commands and tests. We use it to:

  • Scaffold and build websites, then refine them page by page.
  • Write the small serverless functions that connect forms to your CRM.
  • Build custom tools such as intake forms and PDF generators.
  • Run repetitive migrations, like converting a folder of images or checking every internal link.

The speed gain is real, and it shows up as more time spent on your business questions and less on boilerplate. But the output is code like any other. It gets tested, reviewed, and deployed through version control, so every change has a record and can be rolled back.

Content pipelines with deterministic quality gates

Our in-house site generator writes copy for local service sites section by section. A language model writes, but a model does not decide whether the writing is acceptable. Plain code does.

A deterministic gate is a check that gives the same answer every time: word counts against a budget, required headings present, banned phrases absent, no dashes, every internal link pointing at a page that exists, every section from the outline actually in the output. Gates do not get tired and do not have a good day or a bad day.

One of our lessons here: a check that compares two documents built at the same moment by the same pass proves nothing, because both can be wrong in the same way. After two service pages shipped with whole sections missing, we added a verify gate that compares the build against the original outline, which nobody else had touched. The build now refuses to finish if something is absent.

Second-model review

When a section fails a gate and the writer's own retries do not fix it, we do not just ask the same model again. The disputed section goes to models from other labs for an independent opinion, and the objection only clears if they agree. If a call fails or either reviewer holds firm, the section stays flagged for a person. It is a last resort, not the first step, partly because it is the one place per-page API cost shows up.

Human approval gates and scheduled publishing

Some of our sites publish articles on a schedule, twice a week, through a scheduled Claude routine. The routine researches a topic from a keyword queue, writes the article, updates the sitemap, cross-links existing articles and notifies search engines. That works for low-risk, tool-focused content with clear rules.

For client work the rule flips. Drafts wait for a person, and publishing is a deliberate git push of an approved commit. Anything not committed physically cannot go live. For anything regulated, legal, or medical, a human signs off before it ships.

Honest limits

  • We gate everything. A model can sound confident and be wrong. We do not publish unreviewed claims, statistics or testimonials, and we do not invent them. If we have no number, the sentence goes without one.
  • Medical data needs 100%. On a lab-report project for a medical practice, we tried OCR on faxed scans. Accuracy reached 95 to 98 percent, and we judged that not good enough for medical data, because anything short of perfect means a person must recheck every value. Text-based PDFs, which can be read exactly, were handled at 100 percent. The tool also runs fully offline for privacy. We would rather tell a client "this input is not reliable yet" than ship a guess.
  • AI does not replace judgment about your business. It does not know your customers, your margins or your reputation. You and we do.
  • Cost matters. We keep prompts lean and use outside-model review only as a last resort, so a pipeline does not quietly run up a bill.

When Claude is not the right tool

If a task has exact answers (arithmetic, matching IDs, formatting rules), plain code is better and cheaper. If the input is unreliable and the stakes are high, a person should be in the loop. And if you simply need a clean five-page website, the tooling matters less than the strategy behind it: see our small business websites work.

Curious how this would apply to your business? Get in touch.

Ready to bring in more customers?

Tell us about your business and what you need. We'll reply with a plan and a fixed price, usually within one business day.