A business asking AI to do serious work needs somewhere to store what the company knows. Without that store, an AI assistant answers general questions well and company-specific questions poorly. A working setup needs no custom engineering. The work is deciding where each piece of company knowledge belongs and setting the structure up inside tools the business already owns.
AI makers want companies to build the knowledge base inside their own platform, because doing so locks a company into one model. The same lock-in makes a platform the fastest place to begin. Starting there and moving later beats waiting for full independence, which rarely gets built from day one. A team plan, built for several people working from the same information, is the natural starting point. Choosing between Anthropic’s Claude and OpenAI’s Codex counts for less than picking one and starting.
Claude breaks the structure into three parts: connectors, projects and skills. Connectors link an account directly to data a company already keeps in tools such as Notion, Gmail or Slack. The assistant then reaches the original records instead of a separate copy. The first rule is never to duplicate data when it can be avoided, because a second copy drifts out of date the moment the original changes.
Projects suit information tied to one scope. One project might hold everything known about a single client whose history runs deeper than a CRM holds. Another might cover one major initiative. Only one project stays open at a time, so every other piece of knowledge the work needs has to live in a connector or a skill.
A skill is a folder of plain text files, with one required file that carries the instructions and as many reference files as the work needs. Skills started as a way to store instructions, but the format can hold any kind of data a business wants on file. Unlike a project, a single request can draw on several skills at once.
In practice, a company builds one skill per business process worth AI’s help, with its templates and exceptions as reference files. A second set covers internal systems, such as the task management software, the CRM and email. The rule against duplication holds here too, so an email skill that needs CRM details links to the CRM skill instead of repeating them.
