Earlier this year, I wrote about how those who wanted to go further with vibe coding should really move to Claude Code. But that’s only the start of the story.

Turns out, Claude Code is not just a great agent for coding, it’s a great agent for work – if you know how to make it so.

So if you are someone who has already found value from Claude Projects, Artifacts, Research, Connectors and more, this is for you. And if you aren’t confident with those, good news – it doesn’t matter, this guide leapfrogs it all to better solutions anyway.

Strap in, let’s talk Claude Code for work.

1. But why though?

By now, you probably know that LLMs are an incredibly powerful core that can be incredibly dumb without the right context.

Generally, context can be enhanced a number of ways:

• better prompting/ docs – more specific inputs will give more specific outputs
• rounds of reasoning/ thinking – this further enriches the prompt and helps it add its own detail and expectations
• tool use – lets it get even more/ more up to date info, or create useful deterministic scripts or apps or data to reference.

The Claude desktop/ mobile app will use reasoning, and sometimes ask questions, to add as much value as it can with the first two. And it can use tools via connectors (MCP) to reach into your other apps.

But it ultimately lives in a walled garden. It’s “on” your computer, not “in” your computer. Even if you add connectors to let it access your file system, everything is cautious by design, because it’s an app made to be used by your granny for recipes.

Claude Code is AI unleashed.

2. What do you mean?

My Claude Code lives inside a folder that contains all my notes, all my meeting notes, syncs with my to do list app and “system” notes that I ask it to make and maintain. This includes notes about people I interact with a lot, client projects, templates for docs I use a lot, my tone of voice for writing, email or strategy planning.

It has a voice that overwrites the default system prompt so it’s no longer a coding robot but my entrepreneurial cofounder and sparring partner. And I can switch the voice to one designed to specialise in a whole new kind of code that a startup I work with is building. Or to be a chef. Or anything else I choose.

It can connect to just about any tool imaginable, you can build Skills that combine both LLM agents and deterministic bits of code to produce predictable outcomes again and again. And those Skills can be shared with anyone to use in the main Claude app.

(And as if that all wasn’t enough, it can also build literally any tool I want because it’s the best coding AI out there.)

3 Sure but, in practice, what do you mean?

Monday morning I get to my desk, open Claude Code and say “what should I focus on this week?”

It can read my to do list, including priority etc – then it goes to find context from the relevant client folder, including recent meetings, cross references this against my current strategic priorities. It reads my daily notes (if it wants to), to find any recent history or sentiment on tasks/ client projects.

Then it comes back with an analysis of priorities and logic. I’m extending this into monthly/ quarterly planning.

I have a meeting at 10am. I tell it to go and research the person, give it a bunch of context (dictating, currently with Whispr Flow – but MacWhisper is also good) and it finds any detail we need from recent meetings, and sets out some useful notes for me to reference.

I’m going to send follow up email later in the day for that meeting – I ask it: “can you prep a few of the suggestions we discussed to send post-meeting I want you to focus on X Y Z.” It reads my email tone of voice doc, makes a first draft. If it’s rubbish, it at least gives me a starting point to do my own little version – and I can tell it to update my tone of voice with that change.

You get the idea. This is the raw firehose of LLM capability, and the above is just scratching the surface.

(And do you know what, I have a call now so I’m just going to ask Claude Code to review this post before I write the rest…)

4. Yes please, how do I get started?

Getting help from the AI isn’t really the hard bit. The difficulty today is getting all your data in a form that it can easily access.

Here’s some principles to get you started:

Track more stuff

Meetings are the obvious one. You might be using a meeting tool, but you’ll be much happier getting its summaries and full transcript in pure text.

Anyone going hard on codified project management and planning is going to have brilliant info for the AI. I currently sync TickTick (a freemium project planner) to a text file.

If you have chats with the AI about anything significant, ask it to save the file. If there are people you work with often, ask Claude to research them and write a short profile – which can include how they tend to respond to things.

Think about starting each day with a short bit of writing, or a short conversation about what you’re trying to do (or transcribe it out loud.)

Use open file formats, synced to your computer

Yes you can connect these apps to Google Docs, your CRM or wherever else – but this is never going to be as quick or powerful as just storing your data simply. LLMs are a constant context challenge, and simple data formats let them jump in and just pluck the sections they need with speed and ease.

I use Obsidian, a notes tool that saves everything in nearly plain text. Spreadsheets can be exported as .csv.

Consider file history

Some of the best context is the changes you make to the documents you use. In the coding world, having a GitHub repo with commits lets you ask the LLM when a piece of work changed, or roll back to old versions with ease.

This can be applied to the world of “normal” work surprisingly well. I delivered a new Claude Skill to a client recently in the form of a repo, so they could see the new features added in Pull Requests, understand where and how the work was being done.

This would be radical for people who work with clients outside of code – instead of capturing timesheets, you can demonstrate the value and effort put into a deliverable through the clear history of what has been done.

Make templates, specifications and guides

Grab 10 versions of the planning or strategy document you use. Give them to Claude and prompt it to look for patterns and particular characteristics of how you approach the work to write a short guide for someone to follow, and a template of the structure.

Now rinse and repeat for all your common workflows – email tone of voice, reports, weekly catchups or calls (transcripts!)

This technology will create different outputs every time, this is one of the key ways to get more consistent outcomes.

5. Go Claude Go

Once you have a good amount of data in this form, you simply boot up Claude Code in the folder and get going.

Open the terminal, type “claude” and you’re away.

Yes, it’s in the terminal, which can seem intimidating. But what you’re looking at is an area where the AI talks to you, with a line below it where you type your messages.

What’s the difference between that and ChatGPT?

The answer is: you are now working in a folder of growing knowledge and context MADE for AI. You can save a breadcrumb trail of your work defined by you, for future value. You might build a folder structure around different client projects, if that’s how your work lives.

Or you might ask it to build little apps and software for you – an interface for your weekly meetings, or a process to help prioritise what to do next.

These are the first steps in a whole new world of work. Lots of people are trying to sell you on their whizzy AI apps today. But almost all of them create their value sitting almost as close to the model as you are able to with Claude Code.

Until we see some truly breakthrough design and UX ideas, I believe you are better off bootstrapping a simple system with Claude Code. For now, I think it’s maybe the greatest unfair advantage you can build with AI.

CLAUDE WORK WILL RETURN IN PART 2 – output styles, Claude Skills and more…

P.S. I’m partnering with Anthropic to organise a London meetup I call Claude Code Curious on 10 Dec, and again in January. Send me a note if you want to know when the page goes live.