The assistant that does the daily work

What it is

Claude Code is Anthropic's command-line coding agent. Rather than suggesting the next line in an editor, it works on a whole task: it reads files, runs commands, makes changes across a project, checks the result and reports back. You describe the job; it does the legwork; you decide whether the outcome is right.

Here it is used for more than code. A lot of running an estate is operations work — reading logs, checking configuration, writing a script to verify something across every host, drafting a runbook. That is the kind of work it is good at.

Where it fits

• Building. New pages, scripts, configuration and small applications, written to the conventions already in the project.
• Operations. Investigating a fault, reading logs, checking a setting across many hosts, making a change and verifying it from the outside.
• Inventory and audit. Walking a large estate methodically and writing down what is there — tedious work for a person, and exactly where thoroughness matters most.
• Documentation. Runbooks and change records written while the work is fresh, rather than reconstructed a month later.

Where it does not

It is fair to be specific about the limits, because they decide how it should be used.

1. It does not carry accountability. Decisions with consequences — what to delete, what to expose, what to tell a client — stay with a person. It proposes; I approve.
2. It can be confidently wrong. A summary of a system can read perfectly and still contain a claim nobody checked. Anything load-bearing is verified against the real system before it is relied on.
3. It is not where secrets belong. Credentials and sensitive data are kept out of what it is asked to read or write wherever possible, and never pasted into a prompt for convenience.
4. It is not a substitute for knowing the system. It speeds up someone who understands what they are looking at. Without that, it simply produces mistakes faster.

How it is set up here

Claude Code runs on my own machines and works against this estate under my supervision. It has the access a given job needs and asks before anything destructive. Server changes it makes go into version control like any other change, so there is a history and a way back.

It reaches the models through the Anthropic API, and each job's cost is accounted for. That matters more than it sounds. An assistant that quietly runs up a bill is a liability; one whose cost per task is visible can be judged on whether that task was worth doing.

What I would do for a client

Most organisations do not need an AI strategy. They need to know which two or three jobs a tool like this would genuinely take off someone's desk, and what guard rails it needs before it touches anything real. I would start there: find the repetitive, well-understood engineering and operations work, set it up with sensible permissions and review, measure what it costs per job, and be honest if the answer turns out to be "not worth it here".

Questions people ask