Here's exactly
how it works.
We're not hiding anything. This is the full stack, the full process, and the full reasoning behind every decision. You could take this page and try to build it yourself. Most people won't, and that's fine. The point is that you understand what you're buying and why it's built the way it is.
Everything below is running in production. The scores at the bottom of this intro are real.
Most modern websites are slow because they were built on assumptions that made sense in 2012 and nobody questioned them. Database-backed CMS. Plugin-dependent functionality. JavaScript frameworks doing work the server should be doing. CDN layers added to compensate for the bloat.
We inverted every one of those assumptions. Flat files. Server-side PHP. No database round trips. No plugin ecosystem to maintain. The result is a site that's fast not because we optimized it, but because there's nothing to slow it down.
The same philosophy applies to the AI layer. Most AI content tools are generic because they have no other option; they don't know your business. We solve that before we write a single line of content, by building a structured knowledge base specific to you. The Knowledge Stockpile is the product. Everything else is an application of it.
Industry research comes first
Before we talk to a single client, we build a deep industry knowledge base. For a water treatment company, that means the technical details of how salt-based softeners work, what homeowners misunderstand, how to compare treatment options, what drives purchasing decisions, and what questions people ask before they call anyone.
This baseline is built using AI-assisted research combined with real search data. We use AnswerThePublic and similar tools to pull every search term and question pattern around the industry's core topics. These become the FAQ structure, the content calendar, and the gap-detection baseline for every client in that vertical.
Once we have an industry dialed in, onboarding a new client in that vertical is fast. The deep work is already done.
We pull actual search queries, not AI-guessed topics. AnswerThePublic surfaces what real people type, including every "people also ask" variation. Those become FAQs, comparison pages, and content targets.
Industry research goes into a dedicated GitHub repository as .md and .csv files. Markdown for structured content, CSV for keyword data. Version controlled from day one, nothing is ever lost.
The conversation that changes everything
We sit down with the business owner for a recorded conversation. Not a questionnaire. Not a form. A conversation, because humans answer questions completely differently out loud than they do in writing. The typed answer to "what makes your business different" is usually three sentences of marketing language. The spoken answer, given enough time and the right follow-up questions, is usually the actual truth.
The recording gets transcribed and fed into Claude alongside the industry baseline. The model identifies what's unique to this business, what's missing from the owner's account that the industry research says should be there, and where the two diverge in ways that need more information.
We pull every available review from Google, Facebook, Yelp, and anywhere else they exist. Claude runs a pattern analysis against what the owner said. Customers often identify a business's strongest differentiator before the owner does. We find that pattern and make sure it leads the marketing.
The transcript comparison identifies two things: what this business does differently from the industry norm (lead with it), and what topics the industry baseline covers that the owner didn't address (go get that information). Nothing gets assumed. Nothing gets invented.
Building the Knowledge Stockpile
Everything from the industry research, the owner interview, the review analysis, and any existing website content gets organized into a structured repository. This is the Knowledge Stockpile. It lives in a dedicated branch of the client's GitHub repository, formatted in Markdown and JSON so any AI model can read it without guessing or hallucinating.
The structure matters. Markdown for narrative content, definitions, and FAQs. JSON for structured data like service areas, pricing logic, team information, and business policies. The combination means Claude can pull exactly what it needs for a given content task without being given everything at once.
Content generation on a schedule
Claude runs autonomously against the repository on a set schedule: daily, every other day, weekly, whatever the client's growth pace calls for. Each run picks the next title from the content queue, reads the relevant sections of the Knowledge Stockpile, and writes the article.
The constraint is architectural, not instructional. Claude cannot access anything outside the repository. It cannot invent a service area the client doesn't cover. It cannot attribute a differentiator the owner didn't claim. If the information isn't in the stockpile, it doesn't appear in the content. That's not a prompt engineering trick. Those are structural guardrails.
Claude generates Open Graph images as SVGs based on the site's branding guidelines. SVGs get pushed to the server with the content. On first page load, they convert to .webp and cache. Zero manual design work per article.
Each article generates a corresponding social post, pushed to OnlySocial via a custom MCP server. Same knowledge source, same voice, no additional work. The post links back to the article and the article feeds the post.
Content generated by Claude gets reviewed by a second model from a different provider before it touches the live site. Different model, different training, different blind spots. The reviewer checks every article against the Knowledge Stockpile: does anything contradict what the business actually said, does anything appear that has no source in the stockpile, do service area and differentiator claims hold up. Pass merges automatically. Fail flags for human review with the specific violation noted. The goal isn't redundancy. It's that two models from different companies are unlikely to hallucinate the same thing in the same direction.
The website stack
The website itself is flat-file PHP hosted on Hostinger, which has native GitHub integration. When content gets pushed to the main branch, GitHub Actions deploys it automatically. No FTP. No manual uploads. No CMS login.
There is no database. The site reads directly from Markdown and JSON files. PHP handles all the parsing server-side on first load, caches what it can, and gets out of the way. The sitemap, the llms.txt file, and the site index are maintained by lightweight PHP processes that check for gaps on page load and fill them without any scheduled job or external API call.
The feedback loop that makes it smarter over time
Most content systems are one-directional: generate, publish, repeat. This one closes the loop. Every week, the system runs PageSpeed Insights against the homepage and a rotating selection of content pages. Performance, accessibility, SEO, and agentic browsing scores get logged. When something drops, Claude proposes a fix.
Search Console data feeds in on the same weekly cycle. Claude looks at which pages are actually getting traffic, what keywords are driving those visits, and whether the content is optimized for what's working. Adjustments get proposed based on real data, not guesses.
Every fix Claude proposes goes to a branch. It does not touch the live site until it has been reviewed and approved. A weekly spot check takes about ten minutes. The AI identifies and proposes. A person decides. That distinction matters, and it is not going to change.
Weekly automated runs against the homepage and spot-checked content pages. Scores logged, regressions flagged, fixes proposed to a branch for review.
Weekly pull of keyword performance and page-level traffic data. Claude reads what's actually working and optimizes the next content cycle toward it. The system learns from real search behavior, not assumptions.
You've seen how it works.
Want to see it for your business?
We'll build a demo Knowledge Stockpile for your industry and walk through the full system. No slides. No pitch. Just the actual thing running.
Request a Live Demo