Content with Claude Code - a brief how-to
First post in this space. We'll be using it for how-tos and random ruminations.
I’m going to give you a breakneck explanation of how to set up your own marvellous mechanical multiagent content machine
You’re going to be using terminal.
If you’re not comfortable with the geek side of this, ask a friendly dev to babysit you for a few hours. I promise they’ll find it interesting. Otherwise wait for Claude Cowork to filter through - that’s a non-technical wrapper on all this.
Anyway, let’s go.
Get Claude Pro.
Install Claude Code on your machine.
Create a folder with 3 subfolders - legislation, guidance, style.
Get the legislation for your domain, your existing GOV.UK guidance, and all the GOV.UK style material in these folders, in markdown (or whatever format tbh - Claude can convert if you ask it. But PDF is very likely to make it choke - avoid!).
Don’t use any internal gov material - everything will be processed on Anthropic’s servers.
Alongside your folders make a markdown file called CLAUDE.md. It reads this when it starts up so just write in it - tell it its role, your goals, any rules. If that sounds daunting, this’d be fine:
This is an environment for analysing, editing and drafting GOV.UK guidance, based on user needs and the principles explained in the style folder.
You can tell Claude to add more as it understands its role.
Now, let’s get to work.
Fire up Claude. Tell it to look around. Explain your goals in more detail and ask it how it can help, what it can do for you, what it needs to get things done. Tell it to make notes as it goes along - it’ll build up the CLAUDE file to explain itself to itself.
Do some basic tasks. Look for gaps in a guide, find contradictions, find messy language (Sandbox README has examples). Ask it to think about users. Ask it to look at the structure of the local Markdown guidance pages, and then point it at the original URL. It will learn.
Give it multiple personas - editor/drafter, accuracy reviewer, style/2i reviewer. Each should have an MD file describing its role - you can ask Claude to wrte this.
Get them to review content in separate instances and contribute to shared notes. It can also do parallel reviews but that will burn through your tokens.
It’ll keep asking you things. Some of them will be permission to run commands and install things. But sometimes... the first time it questioned me about breaking the 65 character limit on a title and discussed front-loading was perturbing.
But it’s asking you! In the end a human content designer (you) is making the actual decisions.
Golden rules: always verify and keep the human in the loop.