Claude setups

A capable model dropped into a team with no setup behaves like a very quick new starter on their first morning. It does not know where anything lives, which conventions matter, or which commands are safe to run, so somebody spends the day correcting it.

Most of that correction is the same every time. Writing it down once, in the place the model reads first, turns the first morning into every morning’s starting point.

Claude Code for a team

Claude Code works inside a repository. A setup gives it an instruction file at the root that says what the project is and where to look for each kind of task, skills for the jobs the team repeats, and hooks that run checks before work is accepted. Connected tools come in through MCP, so it can read the systems the team already uses rather than asking someone to paste from them.

This site is built that way. Its repository opens with an instruction file that sends an agent to the right document for each task, and a verify step that checks types, house style, and the built pages must pass before anything is committed. Our own notes vault has hooks that check every edit it makes.

Claude for a business

For a team that works in documents and email rather than code, the setup is Claude itself. Projects hold the business’s own documents and standing instructions, so a question about a policy or a price list is answered from the real one. Connectors reach the tools the team already uses, and the defaults are set so that nobody has to explain the business again in each conversation.

Custom agents on Claude

Where a job needs an agent of its own, we build it on Claude’s API or its Agent SDK and wire it into the client’s systems. The same rules apply as in any automation that can change records: a preview before it acts, a clear boundary on what it may touch, and a way to undo a bad run.

What it costs

Claude is paid per seat or per use, on the client’s own account, so the cost belongs to the business and stays visible to it. The setup itself is a written body of instructions, skills, and checks that the team owns and can change.

That written body is the real product. A model improves every few months on its own, while the knowledge of how a particular business works only exists once somebody writes it down.