From prompt to product: building Orbit and Cortex with Claude
Two products, one stack
Over the past year we built two things at Fautons. Cortex is a brand-knowledge server: it holds a company's voice, proof, personas, and policies as structured data any AI agent can query. Orbit is a content engine: a pipeline of agents that turns a brief into an on-brand, proof-backed draft. Orbit reads from Cortex on every run.
Both run on Claude, and both talk to the rest of the stack through MCP, the open protocol Anthropic introduced for connecting AI to data and tools. I am writing this up because the useful lessons were not the ones I expected going in, and most of them transfer to anything you build on Claude.
Lesson 1: structure beats a clever prompt
The first version of Orbit was mostly prompt. A long, careful instruction that tried to carry the brand, the rules, and the task in one go. It demoed well and drifted in production. Two teammates would get different voices for the same brand, because the context lived in whatever each of them had pasted that day.
What fixed it was boring: we moved the brand out of the prompt and into a typed store, which became Cortex. Once voice and proof were data the agent looked up rather than text it had to interpret, the output stopped wandering. I spent weeks trying to prompt my way to consistency. A day of modelling the knowledge did more.
Lesson 2: split the work into stages you can inspect
Asking one model to plan, research, write, and edit in a single pass gives you something competent and a little generic, and when it is wrong you cannot tell which part failed. So Orbit became a pipeline instead: understand, decide, research, generate, humanize, ship. Each stage has one job.
The win is not that it sounds more sophisticated. It is that when a draft comes out weak, I can see whether the research stage missed sources or the humanize stage ignored a voice rule, and fix that one stage. A pipeline you can inspect beats a one-shot you cannot, every time something goes wrong, which is often.
Lesson 3: design for a human in the loop
Neither product is fully autonomous, and that is a choice, not a limitation we are apologising for. Orbit assembles a publish-ready draft; a person still reads it before it ships. Cortex serves knowledge; a person still decides what counts as approved truth. We put the approval points in deliberately, at the moments where a mistake would be expensive.
This is the part people skip when they are excited. The honest framing is that Claude does the heavy, mechanical work and a human keeps judgement over the parts that carry risk. Build the approval step in from the start. Retrofitting trust after something embarrassing has shipped is much harder.
If you're building your own
You do not need our scale to use any of this. Start with the smallest version: write your brand's real rules down as data instead of paragraphs, and split one task into two stages, a draft and a critique, before you reach for anything fancier. Those two moves carry most of the benefit. The building blocks are covered in our guides to multi-agent content engines, giving your AI a company brain with MCP, and building business apps with Claude.
If you want to watch the building and the judgement together, we make something real with Claude every Saturday in AI Pulse, and our hands-on AI training is where teams learn to do it inside sensible guardrails. You can also see the two products these lessons came from: Cortex and Orbit.
Frequently asked questions
Are Orbit and Cortex built with Claude?
Yes. Both run on Claude and connect to the wider stack through MCP, the open protocol for linking AI to data and tools. Cortex is a brand-knowledge server; Orbit is a content engine whose agents read from Cortex on every run.
What is the single most useful lesson from building them?
Structure beats a clever prompt. Modelling knowledge as typed data and splitting work into inspectable stages fixed more than any amount of prompt-tuning. The reliability came from the architecture, not the wording.
Does building on Claude mean removing humans from the work?
No, and we would argue against it. Both products keep a person in the loop at the points where a mistake is costly. Claude does the mechanical work; a human keeps judgement over what ships. Design that approval step in from the start.
Can I apply these lessons without using Orbit or Cortex?
Yes. Write your real rules down as data rather than prose, and split one task into a draft stage and a critique stage. Those two habits transfer to anything you build on Claude, no specific product required.
