Claude Code for Business Operations: It Is All Files

Claude Code is built for development, and that framing hides what it actually is: an agentic harness over files. Strategy documents, emails, meeting reports, agendas, insights, they are all files, and as long as the right MCP connector exists, Claude Code handles them with the same rigor as code. It is the first agentic harness I actually trust to do good work; people talked about agents for a year while the implementations were bad, and this one is not. Here is one boring, real workflow, end to end, to show the shape of it.

The workflow: brain dump to reviewed document

Voice in. I talk to my computer, using Monologue for transcription, and dump a half-formed thought: something about solopreneurs, division of labour, and why an operating system for one-person companies gets more customers as that shift accelerates. Voice matters here: typing makes you brief because typing is work, and talking makes you generous with context. Chaos in is fine, because something structured comes out.

File out. Claude Code knows where insights live in my documents repository, writes the thought into a proper document, and commits when asked. The repo part is the trick most people miss: documents in git means documents get pull requests.

Adversarial review. Then I hand the draft to O3, exposed through a small MCP utility I built, to challenge it. It came back with real objections: cherry-picked examples, sectors where a small team cannot compete, the fact that AI helps large companies too, a bias toward low-capital low-regulation sectors, and the observation that all of it builds on top of old-money institutions. My naive brainstorm was more nuanced after the beating. I use O3 for exactly two things: very specific code changes when Opus fails at them, and critical takes, because it is very good at being unimpressed.

PR as review gate. The refined document goes into a pull request rather than straight to master, even though it is just prose and nothing can break. The PR view shows what changed, and inline comments become instructions: mark a line, mention @claude, and the GitHub Claude agent makes the change in the PR, after which the automatic review that ships with the GitHub app critiques its own commit. Document review with the entire code-review toolchain, for free.

Assistant, agentic, or background: which mode when

Three modes, one decision. Assistant mode is the back-and-forth above, right for work you are still shaping. Agentic mode chains the steps without you; the challenge-and-refine loop could become a single command once the process is stable. Background agents, like the GitHub one, fit self-contained tasks: no worktree setup, no merge dance, just a ticket. The catch is that a background agent's only context is the ticket, so the ticket has to be very good.

On MCPs, since connectors are what make all of this reach your tools: prefer official ones, or wrap an API yourself; it is not rocket science, auth is the only genuinely annoying part, and I have Claude Code edit my own MCPs when they need changes. Be careful with third-party MCP marketplaces.

The actual point: you cannot outsource this understanding

Right now this is less about automation than about learning where each tool fits: harness in the driver's seat, background agent, vibe-coding tool, each has a place. Once you have run a process manually and codified it, you automate one use case after the next; that manual-first loop is the same everywhere. Most people try to skip to "AI can do everything," often on the advice of people who have never done any of it themselves.

The old delegation rule applies: only delegate tasks you understand. Learn the tool yourself, find its strengths and weaknesses, then hand the process off, to a person or an agent. The leverage on that understanding is higher than it has ever been, and nobody can take it away from you once you have it.