A well-built Claude Code workspace turns AI from a tool you visit into an operational system that already understands your business. Too many founders still use Claude like a chatbot: paste in context, ask a question, copy the answer, close the tab, then repeat the whole process tomorrow. Nothing is retained, nothing connects, and none of the effort compounds.
The alternative is a structured workspace that holds your business context, live data, daily briefs, and reusable commands in one place. Claude begins each task with a working understanding of what you sell, who your team is, what matters this week, and what the numbers are saying.
This post shows you the practical setup: the folders to create, the files worth building, and the order to put them together so your workspace becomes genuinely useful—not just another place to store information.
Key Takeaways
- A Claude Code workspace turns AI from a one-off chatbot into a persistent operating system for your business.
- The essential structure includes
context/,data/,scripts/,outputs/,reference/, and a rootCLAUDE.mdfile. - Context files give Claude the business knowledge it needs to produce relevant, non-generic work without repeated explanations.
- Start with five core context files: business, team, strategy, clients, and history.
- A simple data layer, such as a SQLite database, lets Claude analyse current CRM, revenue, project, or accounting information.
- A daily brief can replace time spent checking dashboards and gathering updates by delivering the most important numbers, risks, and priorities in one place.
- Reusable slash commands turn recurring work—such as follow-ups, client updates, and weekly reviews—into repeatable workflows.
- Keep the workspace lean. Store only essential information in always-loaded files and keep detailed documents in
reference/for use when needed. - Build gradually: create the structure, add context, connect one data source, create one useful command, and improve it through real use.
Why a Workspace Beats a Chatbot
The chatbot model of AI use is fundamentally limited. Every conversation starts at zero. Every prompt re-explains your situation. Every answer is generic because the AI has no memory of your business between sessions.
A workspace flips this. Claude Code reads from local files before it responds. Those files contain your strategy, your team roles, your client list, and your latest metrics. The AI is briefed before you type the first word. It’s the difference between calling a temp agency every morning and having a chief of staff who’s been with you for three years.
The practical result: you stop pasting context. You stop re-explaining. You type one command and Claude produces work that’s already informed by everything your business knows. Not generic output. Output shaped by your specific context, your actual numbers, your current priorities.
This is what Liam Ottley calls the intelligence layer. It’s not a prompt library. It’s not a SaaS subscription. It’s an operating system that wraps around your business and starts thinking for it.
The Folder Structure That Actually Works
Every Claude Code workspace has the same core shape. You can add to it, but these folders are non-negotiable.
context/ holds the business brain. Who you are, what you sell, who’s on the team, what the current strategy is. This is the first thing Claude reads in every conversation. If your context folder is thin, every answer will be generic. If it’s thick, every answer is sharp.
data/ holds your numbers. Exports from your CRM, accounting software, and analytics tools. A single SQLite database works well for most businesses. The AI reads from this when you ask anything number-related.
scripts/ holds your automations. Python scripts, shell commands, anything that collects data or runs a workflow. This folder grows over time as you automate more tasks.
outputs/ holds the deliverables. Drafts, reports, briefs, content. Anything Claude produces on your behalf lives here so you can review before it goes anywhere.
reference/ holds the on-demand docs. Templates, prompts, and detailed files that only get loaded when relevant. Keeps your always-on context lean.
CLAUDE.md sits at the root. This is the file Claude reads first, every time. It points to everything else. Keep it terse. Push detail into reference/ files that load on demand.
That’s the skeleton. Five folders and one file. Most businesses can run on exactly this structure for months before needing anything else.
Layer 1: The Context Files
This is where 90% of founders get stuck. They know they should document their business. They’ve tried. The docs get written, filed, and never opened again.
A context file is different because it’s not documentation. It’s a briefing for an AI that will actually use it. Every day. In every conversation. The test isn’t “would a new hire read this?” The test is “can Claude answer strategic questions about my business after reading this?”
Start with five files:
business.md covers what you do, who you serve, what you sell, and at what prices. One page. Dense, specific, no fluff. If a stranger reads this file, they should know what your business is in 90 seconds.
team.md lists every person (employees and regular contractors), their role, what they own, what they’re working on right now, and what they’re not good at. The last part matters. Claude needs to know not to route certain work to certain people.
strategy.md holds the current focus. What are you actually building this quarter? What are you deliberately not doing? What’s the number you’re trying to move? This file should change every few months.
clients.md covers your current client list (or customer segments), with notes on each one. Names, status, last contact, outstanding items. If you have 200 clients, use categories instead of individuals.
history.md captures the origin story and recent inflexion points. Why does the business exist? What has already been tried? What worked, what didn’t? This stops Claude from suggesting things you already tested three years ago.
Write these files in plain English. No structured templates. Claude reads natural language better than it reads forms. Just write like you’re briefing a new senior executive on their first day, except this executive never forgets and is available 24 hours a day.
Spend two hours on this. Not two weeks. The first version is always rough. You’ll improve it every time you notice Claude missing context in an answer.
Layer 2: The Data Layer
Context tells Claude what your business is. Data tells Claude what’s happening right now.
The goal: one place where all your numbers live, refreshed automatically, readable by the AI without you having to export anything manually.
For most businesses, a SQLite file (data/data.db) is sufficient. It’s a single file, no server to run; Claude can query it directly with Python. You don’t need a data warehouse. You don’t need BigQuery. You need one file that holds the numbers that actually drive decisions.
What to connect first: pick the one source that answers “how are we doing?” For most service businesses, that’s the CRM. For product businesses, it’s often accounting. For agencies, it’s usually a combination of revenue and project hours. Don’t try to connect everything at once. Pick one. Get it working. Move on.
The mechanism is simple. A Python script runs on a schedule (daily or hourly), pulls data from the source via its API, and writes it into your SQLite file. That’s it. No fancy ETL. No cloud infrastructure. Just a script that runs at 2 am and a database file that’s fresh by the time you wake up.
Once you have one source connected, add a second. Then a third. Six months in, you might have eight sources flowing into the same file. The power is cumulative. Each new source makes every analysis Claude does more complete.
There’s detail on how to architect this in our AI business intelligence system guide and how an AI operating system connects data layers.
Layer 3: The Daily Brief
This is the moment most founders realise the workspace isn’t just useful, it’s transformative.
The daily brief is an automated report that runs overnight, reads your context files, queries your data layer, pulls in any meeting transcripts or messages from the previous day, and synthesises everything into a 5-minute read. It lands on your phone via Telegram or email before you’re out of bed.
What a good brief contains:
– Key revenue and pipeline numbers from yesterday
– Anything that changed significantly (up or down)
– Team updates pulled from meeting transcripts and messages
– Risks flagged across the business
– Strategic priorities for the day, informed by what happened yesterday
The mechanism: a scheduled Python script, a prompt that knows how to read your workspace, and a delivery mechanism (Telegram bot, email, whatever you prefer). Total build time: a few hours if you’re starting from scratch, 15 minutes if you borrow an existing module.
The reason this matters isn’t the information itself. It’s what you stop doing. You stop checking six dashboards. You stop sitting in meetings just to stay informed. You stop spending 90 minutes every morning piecing together what happened while you slept. You open your phone, read the brief, and you’re fully up to date before your first coffee.
This is also the point where away-from-desk autonomy starts climbing. You can read the brief from a cafe, a plane, or on holiday. Make two decisions from your phone, put it away, and get on with your life. The business runs whether or not you’re sitting at a desk.
Layer 4: Automations and Reusable Commands
Context, data, and briefs are the foundation. This is where you start crossing tasks off permanently.
A reusable command is a prompt that lives as a file, takes input, reads the relevant parts of your workspace, and produces a deliverable. You trigger it with a single line. Claude does the rest.
Examples from a real workspace:
/draft-reply— reads the email thread, pulls context on the sender from your CRM, drafts a reply in your voice/weekly-review— summarises what happened across the business in the last seven days/client-update— generates a status report for any client by name, pulling from meeting transcripts and project data/follow-up— identifies leads who haven’t been contacted in X days and drafts personalised follow-up messages
Each command is a markdown file in .claude/commands/ describing what it does and what inputs it needs. You can build one in 15 minutes. You can have fifty after a month of real use.
The trick is noticing the pattern. Whenever you find yourself doing the same type of work twice, stop. Write it up as a command. Next time you need it, you type one line instead of doing the work again from scratch.
This is how the Task Automation percentage climbs. Start at zero. Hit 20%, and you feel the difference. Target 60-70% within six months. For a deeper walkthrough of what to automate first, see this breakdown on AI workflow automation.
Layer 5: Keep It Lean
The biggest mistake founders make after setting up a Claude Code workspace is adding too much. Every file you load into context costs tokens. Every token costs money and slows responses. More context isn’t better context.
Two rules:
Always-loaded context stays minimal. Your CLAUDE.md and the core context files should be short. Under a few thousand words total. Anything longer gets moved to reference/ and loaded only when the task calls for it.
Subagents handle research. When Claude needs to read a big file or search through 50 documents, use a subagent. The subagent reads everything, returns a summary, and the main context only sees the summary. Your workspace stays fast and cheap.
A well-run workspace costs around $20-50 a month in API fees for a solo founder using it heavily. If yours is costing more, something is wrong. Usually it’s context bloat. Fix the bloat, the cost drops, the speed improves, the answers get sharper. Anthropic’s own guidance on context engineering covers this in depth.
What It Actually Looks Like Running
Here’s what my own day looks like with a workspace built across these five layers.
6:45 am. The brief arrives on my phone. Revenue numbers from yesterday, pipeline changes, meeting summaries from calls I didn’t attend, anything flagged as a risk, priorities for today.
7:10 am. I read it over coffee. Reply to two items that need a decision. Flag three items for later. Put the phone down.
9:00 am. First call. Afterwards, I run /post-call-recap, and Claude drafts the follow-up email using the call transcript, the prospect’s context from the CRM, and the voice I always use. I review, adjust two sentences, and send.
11:00 am. I want to check how a specific client project is tracking. I ask Claude. It reads the project notes, the last three meeting transcripts, the budget in the data layer, and gives me a two-paragraph answer with the specific numbers I need.
3:00 pm. New content idea. I run /brainstorm , and Claude produces five angles informed by my actual ICA, current strategy, and the content bible that lives in the reference folder. I pick one.
5:30 pm. Close the laptop. Tomorrow morning, the brief arrives again.
That’s the shape of it. Nothing spectacular in any single moment. What’s different is the cumulative effect. Context doesn’t have to be re-established. Data doesn’t have to be pulled manually. Recurring tasks don’t consume real thinking time. The business has its own brain, and I’m free to work on strategy and growth instead of firefighting.
Where to Start This Week
If you’re building your first Claude Code workspace, do this in order:
- Install Claude Code and create the folder structure. Thirty minutes.
- Write the five context files. Two hours. Don’t overthink it. Rough drafts are fine.
- Pick your one most important data source. Connect it via a Python script. Two to four hours.
- Build the daily brief command. Two hours, or 15 minutes if you borrow an existing module.
- Live with it for a week. Notice what feels rough. Improve those files.
That’s the first layer of a working system. You’ll have more than most businesses will have in the next five years. From there, it compounds. Each command you add, each data source you connect, each context file you refine makes the whole system more capable.
The trap to avoid: don’t try to build everything at once. Don’t try to automate the most complex thing first. Don’t spend three weeks designing the perfect folder structure before writing a single file. Build one layer, use it for a week, build the next.
The Bigger Picture
Setting up this system is the bridge between having a business that relies on your constant input and one that can operate with its own degree of intelligence. When your knowledge, data, and processes are locked in your head, the business can only grow as fast as you can move.
The transition to a Claude Code workspace changes that by giving the AI a persistent, structured environment to work from. By organising your context and commands into a dedicated folder structure, you ensure that every task the system performs is informed by your actual business goals rather than generic assumptions.
This shift allows you to move beyond being the bottleneck for every routine decision or administrative chore. Once you have a working system in place, you can finally focus your energy on high-level strategy and growth while the operational tasks continue to run predictably in the background. For more on how the full system fits together, see our guide to building an AI operating system for business.
If you’d like to map this out for your specific business, book a 30-minute Discovery Call. I’ll walk you through what AI could realistically take off your plate, how to roll it out properly at your size, and whether there’s a fit. No pitch, no obligation.
Frequently Asked Questions
What is a Claude Code workspace?
A Claude Code workspace is a structured folder containing the business context, data, scripts, reference documents, and reusable commands Claude needs to do informed work for your business.
Why is a workspace better than using a chatbot?
A chatbot starts each conversation with limited context. A workspace gives Claude access to persistent information about your business, clients, team, priorities, and data, so it can provide more accurate and useful outputs from the start.
What should I put in CLAUDE.md?
Keep CLAUDE.md short and practical. It should introduce your business, explain where key files live, outline important operating rules, and point Claude to the most relevant context and reference materials.
Do I need a complex database?
No. For many small businesses, a single SQLite database is enough to store and query key information from your CRM, accounting platform, analytics tools, or project systems.
What is a daily brief?
A daily brief is an automated summary generated from your business context and current data. It can highlight revenue, pipeline changes, team updates, risks, and priorities before you start your day.
What should I automate first?
Start with repetitive, structured work that follows a clear process, such as drafting follow-ups, preparing client updates, summarising meetings, or producing weekly reviews. Avoid high-judgment decisions at the beginning.
How do reusable commands work?
Reusable commands are saved instructions that tell Claude what to do, what files to read, and what output to create. Instead of retyping a long prompt, you trigger the workflow with a short command such as /follow-up.
How can I avoid making the workspace too expensive or slow?
Keep always-loaded context concise, move detailed material into reference/, and use subagents or targeted file loading for large research tasks. More context is not always better—relevant context is.
About Octavius
Titus Mulquiney is the founder of Octavius AI, where he builds AI brains and AI workforces for founder-led businesses stuck running everything out of their own head. Twenty years in marketing, ex-Sony product manager, ex-GM Zeal NZ. Based in Auckland, working with operators across NZ, Australia, and the US. Connect on LinkedIn.