Why We Run Our Agency on Claude Code

I run Skein on Claude Code because agency work is files: briefs, research, methods, drafts and reports. Claude Code works through a folder of them, follows the written method for each job, and asks before it changes anything. The business owns every file, the method is shared, and a person approves what leaves.

Why does Skein run its agency on Claude Code?

Because an agency's work is a body of files, and Claude Code is built to work across one. It reads the brief, the client's facts and the method, does the job and writes a draft for review. Everything stays in plain files the business owns, versioned, so any change can be seen and reversed.

Skein runs its own business on a Company OS, and every report, plan and piece of content it delivers goes through it. That system needed a home that could hold knowledge, methods and outputs together. A chat window holds a conversation. A repository holds a business.

Anthropic describes Claude Code as an agent that plans, asks clarifying questions and works on tasks that can run for hours or days. I mostly use it for something less dramatic: running the same written method, reliably, against a different client's material each time.

Why files instead of a chat window?

Files outlast the tool. A method written as a plain file can be read by Claude today and by another assistant tomorrow. It can be versioned, reviewed and handed to a new colleague. Knowledge kept inside one chat tool's settings belongs to that tool, and often to one person's account.

The difference shows up in three places:

Ownership is the one clients care about most. A Company OS should keep working if Skein leaves, so knowledge, agents and workflows live in files and accounts the client owns. The guide to Claude Code for non-developers covers what that looks like day to day.

How do skills keep the method the same across clients?

Each repeatable job is written once as a skill: the steps, the checks and an example of good output. Every client's work runs through the same skill, reading that client's own facts. So the method improves in one place and every client benefits, while one client's details never leak into another's work.

That separation is the design rule I care about most. How an audit is run is method, and it lives in a shared skill. What is true about a particular client is their knowledge, and it lives in their own workspace. Mixing the two makes both harder to keep right.

Skills are also portable. According to the Agent Skills site, a skill built once can be used "across any skills-compatible agent". So the method is not tied to one vendor either.

How does Claude Code keep control with a person?

Three layers. Claude Code asks before it edits files or runs commands. Its settings let you allow, ask about or deny specific tools. And hooks run your own code at fixed points, so a rule holds whatever the model decides. On top of that, a named person approves anything published, sent or spent.

The distinction between guidance and enforcement matters. Claude Code's documentation says instruction files are treated "as context, not enforced configuration". Hooks are different. According to the hooks guide, they give "deterministic control: certain actions always happen rather than relying on the LLM to choose to run them."

So the rules that matter most are held in code. Every connection to a client system starts read only, and write access is granted one tool and one action at a time. Nothing is published, sent or spent without a named person approving it. The write-up of lessons from our Company OS describes where those approval points sit.

What are the trade-offs?

It asks more of the people using it. Claude Code works in a terminal or editor, against folders and plain text, and that is unfamiliar to many teams. Setting up skills, permissions and connections takes real thought. And because it can run commands and reach tools, careless setup carries more risk than a chat window.

There are softer costs too. Context has limits, so large jobs need breaking into steps. Plain files need tidy habits, or the folder becomes as messy as any shared drive. And somebody has to own the setup, keep it current and review what changes.

None of these outweighs the benefits for an agency. They might for a business where nobody is comfortable in a terminal.

Would Claude Code suit your business?

It suits a business with a technically comfortable owner for its AI and repeatable, file-heavy work: reports, proposals, research, content. It is a poor first tool for most staff, who will do better with a chat assistant and shared projects. Many teams end up with both, reading the same shared files underneath.

The pricing makes trying it cheap. Anthropic's pricing page lists Claude Code as included in the Pro plan, as checked in October 2026. Start with one job you already do from files, such as a weekly report, and compare the result with your own.

Skein builds Company OS setups that run from files the business owns, with method in shared skills and a person approving every output; the AI engineering page shows how work flows through them. That suits a growing B2B team with someone ready to own its AI setup. A team without that person is better served by shared projects inside the chat app.

Frequently asked questions

What is Claude Code?

Claude Code is Anthropic's agent for working inside a folder of files. It reads files, follows written instructions, runs commands and makes changes, asking permission before it edits files or runs commands. It was built for software, but it works on any body of files, including business documents and methods.

Can an agency run on Claude Code?

Yes, if its work can live as files and someone is comfortable with folders and a terminal. Knowledge, methods and outputs sit in a repository the business owns. Claude Code runs the written method for each job, and a person reviews and approves every output before it reaches a client.

Is Claude Code only for developers?

It was designed for developers, but much of its value is general: reading many files, following a written method and producing a draft. Non-developers who are comfortable with plain text and a terminal or editor can use it well. Most staff will still prefer a chat assistant.

How do you keep Claude Code safe with client work?

Use its permission prompts, deny or allow specific tools in settings, and put hard rules in hooks that run your own code. Connect client systems read only first and add write access one action at a time. Keep each client's details in their own workspace, and approve every output before it leaves.

Do you need Claude Code to use skills?

No. Skills work in the Claude apps and through the API as well as in Claude Code, and the open standard lets other compatible assistants load them too. Claude Code suits teams that want skills, knowledge and outputs together in one versioned folder the business owns.

Related reading