AI Workspace
The AI Workspace: Bring your context, business systems, and skills together so AI works on your real data, securely.
You were told to use AI. So you tried it.
It drafts an email in seconds, summarizes a long document, answers almost anything you ask. In a demo, it looks like the future.
![](data:image/svg+xml,<svg display="block" id="4166636898" role="presentation" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M 18 11.55 L 18 9.09 C 18 8.231 18 7.803 17.829 7.416 C 17.658 7.031 17.333 6.726 16.682 6.119 L 11.354 1.147 C 10.792 0.623 10.512 0.361 10.163 0.206 C 10.091 0.173 10.017 0.145 9.942 0.12 C 9.578 0 9.181 0 8.388 0 C 4.737 0 2.912 0 1.675 0.93 C 1.425 1.118 1.198 1.33 0.997 1.563 C 0 2.719 0 4.422 0 7.829 L 0 11.55 M 10.125 0.525 L 10.125 1.05 C 10.125 4.019 10.125 5.505 11.114 6.427 C 12.102 7.35 13.693 7.35 16.875 7.35 L 17.437 7.35 M 17.719 14.7 L 14.906 14.7 C 14.285 14.7 13.781 15.17 13.781 15.75 L 13.781 17.85 M 13.781 17.85 L 13.781 21 M 13.781 17.85 L 17.156 17.85 M 0.281 21 L 0.281 18.375 M 0.281 18.375 L 0.281 14.7 L 2.25 14.7 C 3.337 14.7 4.219 15.523 4.219 16.537 C 4.219 17.552 3.337 18.375 2.25 18.375 Z M 7.031 14.7 L 8.719 14.7 C 9.961 14.7 10.969 15.64 10.969 16.8 L 10.969 18.9 C 10.969 20.06 9.961 21 8.719 21 L 7.031 21 Z" fill="transparent" height="21.00000027156895px" id="aYpKLXDcH" stroke-dasharray="" stroke-linecap="round" stroke-linejoin="round" stroke-width="1.61" stroke="rgb(0, 0, 0)" transform="translate(3.085 1)" width="18px"/></svg>)
But come Friday, the actual work is not much further along. The proposal still took the same amount of time. The reporting still lives in the same spreadsheets. The decisions still required the same manual assembly of information from the same scattered systems. The cool demos never quite became done work.
The model is capable. It just does not know your business.
WITHOUT A WORKSPACE
Paste your context every session
Re-explain what you need
Carry the answer back out by hand
Start over tomorrow
WITH A WORKSPACE
Context persists and compounds
Tools are connected, not copy-pasted
Workflows run the same way every tim
To get the most value from AI you have to stop treating it as a clever thing you visit and start building a workspace around the workflow(s) you want to solve. This is a single place where context, tools, repeatable processes, and data live, so the model is finally working on the real thing instead of a demo.
We have been building these workspaces for our own team and for partners, and the pattern holds across industries, team sizes, and technical comfort levels. This guide shows what that workspace is, how to build a personal one you can start this week, and where it goes when a whole team works this way.
01
What a workspace is
An AI Workspace is a structured place that holds three things: your context (what the AI needs to know about your business and your work), your connections (the tools and data it can reach), and your skills (the repeatable workflows you have taught it).
The AI tool that operates inside it, whether that is Claude, ChatGPT/Codex, or something else, is your choice. And it can change.
A chatbot is a doorway you walk through and leave. A workspace is a room set up for your work that stays set up.
Your Tools and Data
Calendar
Task Manager
CRM
Email/Files
Custom App
AI Workspace
context / skills / connections
Reports
Documents
Decisions
Automations
outputs
Two commitments that make this work
Your systems stay the source of truth. The workspace does not copy your data into a separate place. The AI reaches into your systems with the same permissions you already have, so it only ever sees what you could already see. Nothing new is opened up.
You bring the AI tool you already use. The workspace is not built for one provider. It works with whatever AI tool fits the job, and it continues to work when that tool changes. Your context, your connections, and your workflows stay with you, not inside one company's app.
This is the spine of the approach. The workspace is yours. The AI provider is rented. That is what makes switching or mixing tools cheap instead of catastrophic, and it is what keeps you from building your most valuable workflows inside a walled garden you cannot leave.
This matters more than it sounds. AI tool pricing is changing fast. Providers are raising prices, shifting plans, and restructuring what you get at each tier. If your workflows, context, and skills live inside one provider's app, a price change becomes a hostage situation. If they live in a workspace you own, a price change is just a reason to swap the tool. The workspace stays. The provider is a choice you revisit.
Key terms
Skill. A structured set of instructions that defines a repeatable workflow toward a specific goal. Think of it like a playbook combined with a recipe: the playbook gives you the goal and the strategy, the recipe gives you the steps and the ingredients. The AI follows the workflow and exercises judgment within it.
MCP (Model Context Protocol). An open standard that lets AI tools talk to your systems through a controlled interface. Think of it as a menu: the system publishes what the AI is allowed to ask for, based on your permissions. The AI reads the menu and calls only what it needs. Everything else stays private. Unlike an export, which is stale the moment it leaves, MCP is a live connection.
CLI and API connections. Not every tool speaks MCP. The workspace can also run commands in a terminal the way a person would (CLI) or call a system's API directly. Both are often wrapped in a skill to make them easier and safer, and depending on the case, either can be the better choice.
02
Why it compounds
In a chatbot, every good result evaporates when you close the tab. In a workspace, the work you do in a session lives on after it ends. The context you assembled, the prompt that worked, the output you shaped. Key decisions get tracked, so the reasoning behind your work is recorded, not lost to a closed window. Knowledge is encoded in a format the AI can read progressively, pulling in the right piece for the task at hand, so it gets more valuable over time.
What about Projects in Claude or ChatGPT?
Most AI tools now offer a "Projects" feature where you can upload documents and have conversations that draw on them. That is a real step up from a blank chat window. But it is still not a workspace. Your context lives inside that provider's app, locked to their format. There are no live connections to your systems. No skills that run a repeatable workflow. And if the provider changes pricing or you want to try a different tool, you start over. A workspace keeps all of that in files you own.
What that looks like in practice
Instead of exporting from each system and stitching spreadsheets by hand, the workspace pulls from all of them and builds the report. Define it once, rerun it anytime. Instead of re-teaching the AI who you are and how your business runs every session, that context lives in the workspace. It already knows. Instead of ferrying data between tools and AI windows by hand, the systems are connected and the AI works where your data already lives. For the things you do the same way every time, turn them into something that just runs: consistent, fast, and without paying the AI to figure it out from scratch each time. Every one of these is the same move. What was ad hoc and disposable becomes built and owned. That is what a workspace buys that a chatbot cannot.
03
Build your own
A workspace is a work surface, not an app.
Workspace is a work surface, not an app. At the end of the day it is a collection of files, but the power is in what those files hold and what they connect to. Because it is plain files in a project, it is portable, not trapped in one vendor's app. You can back it up, share it, and open it with whatever tool you choose.
This distinction matters more than it sounds. The same coding tools that power a workspace (Claude Code, Codex) can also be used to build custom applications from scratch. This can be worth exploring. But building software carries real risks: security vulnerabilities, bugs, deployment, ongoing maintenance, and the responsibility of putting something into the hands of others. A workspace is not that. You are not shipping code into production. You are building structure, context, instructions, connections, and workflows, that makes the AI's inference more accurate and more useful within your work. You might write a small script or automation along the way, but the core of what you are building is a curated environment, not a software product.
That makes this far more accessible to business users than trying to build a custom application. And it carries a fundamentally different risk profile. If someone on your team does want to use these tools to build software, that is a valid path, but it is one worth doing with guidance, proper review, and an understanding of what they are taking on. And when custom software is the right answer, it layers back into the workspace as another connected source rather than replacing it.
The anatomy
Instructions are the map. A file called AGENTS.md tells any AI tool how your workspace works. It is an open standard: the same file works across Claude Code, OpenAI's Codex, GitHub Copilot, and others. You write it once. It acts as a map, pointing to the right context rather than trying to contain everything. For Claude specifically, a thin CLAUDE.md file points to AGENTS.md, keeping instructions in one shared place.
Context is what persists. The facts and decisions that outlive any single session. How you work. How your business runs. What was decided and why. Organized into areas the AI reads as it needs them. The first week, it asks you basic questions about your setup. A month in, it already knows your conventions and decisions and catches things you would miss. The context compounds.
Skills are the workflows. Repeatable workflows you have taught the AI: a morning brief, weekly review, a meeting debrief. Each one runs the same way every time and improves as you refine it.
Connections are your tools and data. Your calendar, your task manager, your CRM, wired in over MCP so the AI can see your real schedule, your real to-dos, your real data. For tools that do not speak MCP yet, a CLI call or an API wrapper, often packaged as a skill, does the job.
How you operate it
You work in the workspace through an AI tool. The best tools for this are labeled "coding" tools: Claude Code, OpenAI's Codex, Google's Gemini CLI. The name is misleading. Both Anthropic1 and OpenAI2 are positioning these for business work, not just development. They power a workspace because they can read and write files, connect to your systems over MCP, and follow structured instructions.
Most of these tools have desktop apps in addition to a terminal UI. Because the workspace is just files built on open standards, you can switch tools without rebuilding anything. If you are already paying for an AI tool (ChatGPT Plus, Claude Pro), you likely already have access to the workspace-capable version. The workspace approach does not add a new subscription. It changes how you use the one you have. Start simple, with context and a few connections. As you learn what works, add skills, connect more systems, and let the workspace evolve alongside the way you work.
Try it: build a personal workspace
Start with the workspace closest to you: one that runs your own day. It tends to hold four things:
Context that makes it yours
Your rhythm, encoded as skills
The tools you actually use
A place where decisions and outputs accumulate instead of evaporating
07
Where to start
AI is not a chatbot you visit. It is a system you build around the workflow you are looking to solve. That might be managing your life, running your business, or serving your team and customers. The personal workspace is something you can start this week. Paste the starter prompts, answer a few questions, and you have a work surface that grows with you. The team and organizational build is where the value compounds, and where the questions get hard enough to be worth doing with a partner who has built it before.
04
Skills: your expertise, repeatable
A skill is a repeatable workflow you have taught the AI: a clear goal, structured steps, and the supporting context that makes the output consistent. At work, the stakes go up. Consider a team that generates client proposals by pulling from a CRM, past project history, and a pricing model. Without a skill, someone assembles that by hand every time. With a skill, the AI gathers the sources, drafts the document, and presents it for the person to review, refine, and approve. The skill encodes not just the steps but the judgment: which history is relevant, how to frame the pricing, what tone fits the relationship. The human stays in the loop for the decisions that matter. The assembly work disappears.
Skills are portable A skill is more than a prompt. It includes structured instructions, reference material, and sometimes scripts, built on an open standard. And because it is portable, it is not locked to you. Hand it to a teammate and they get your proven approach on day one, the steps that reliably produce a good result, so they do not have to figure it out from scratch.
Why this is different from buying software Generic software gives everyone the same generic features. A skill captures the way your team does something well, and it can be improved over time. The skill improves through use. A shared skill library means the whole team runs the best version of a workflow, not each person's ad-hoc guess. Go deeper Building and structuring skills well is its own craft. See the AI Skills guide for the full breakdown on: when to build one, how to structure it, and how to share it with a team.
05
Risks and control
Once AI can reach your tools and data, the question stops being "what can it do" and becomes a better one: what should it do on its own, and what stays under your hand? That is your call to make.
Your workspace, not your whole computer An important distinction. These tools can be confined to work only with the files in your workspace, not your entire machine. You set a trust boundary, and the AI works freely inside it and asks before stepping outside. Know whether you are giving it access to your whole computer or just this work surface. This is different from features like ChatGPT's computer use or Claude's desktop automation, which interact with your full screen and carry higher risk. A workspace operates within a defined boundary, not across your whole machine.
Start read-only, expand deliberately MCP connections work best as read-only to start. Low risk, high value. The AI reads your calendar, pulls your project data, searches your CRM. It does not change anything. When you are ready for writes, you can make every write a permission ask, so the AI proposes an action and you approve it. Autonomy is something you grant deliberately, not the default. Be careful what you allow an AI to do without asking you first.
Authenticate as you, not as a system You sign in as your own user, so the AI inherits exactly your permissions, never more. Not a system-wide key that hands it the keys to everything. Your systems stay the source of truth. Nothing is copied to a separate place. Every call is logged: who, what, when. When a tool does not offer a read-only mode, you can put a filter in front of it that blocks the risky actions before the AI ever sees them. The AI literally cannot call a function it was never offered.
Where does my data go? Know whether a tool trains on your data. Prefer business or enterprise plans where your data stays private. And note the difference: connecting to your systems through a controlled interface keeps you in charge of what is exposed, unlike pasting sensitive information into a public chat box. Go deeper Choosing the right tools, evaluating their data practices, and building an organizational AI policy are covered in the AI Tooling guide.
Governance belongs with the business The tools that make a workspace powerful are the same tools that make leadership nervous. Non-technical staff can now connect systems, build lightweight automations, and spin up AIpowered workflows on their own. That is a real capability, and it is a real risk if it happens before the organization has clear guardrails around data access, security, review, and ownership. "Use AI responsibly" is not a policy. It gives people no help when they are deciding whether to connect a system, paste in a customer list, or let the AI run a report on its own. The workspace model addresses this directly: the guardrails are built into the workspace itself. What is connected, what the AI can read vs. write, what runs automatically and what requires approval, these are decisions you make once and encode for the team, not choices each person makes in the moment. Getting governance right is where bringing in help starts to pay off. The questions are specific to your data, your team, and your risk profile, and the answers need to be in place before the workspace scales.
06
From personal to team
If you built the personal workspace from the previous sections, you already understand how the team version works. The anatomy is the same: instructions, context, skills, connections. The personal workspace is not a stepping stone you leave behind. It is the same pattern at a different scope. What you learned building your own morning brief and connecting your own calendar is exactly what it looks like when a team shares context, ships skills to each other, and connects the systems everyone relies on.
What changes at team scale The difference is what gets shared. Instructions become a shared map of how the team works. Context captures decisions and institutional knowledge. Skills encode how the team does the work well. Connections expand to reach the systems the business runs on. If your team already works in a custom application, a CRM, or reporting tools, those become sources the workspace reaches directly. If the team's data lives across spreadsheets, documents, and email, the workspace can be the place that starts bringing it together. Either way, the AI stops working from what someone pasted in and starts working from what the business actually knows.
Cost and portability at scale Because nothing is locked to one provider, you can mix tools across the team. One person prefers Claude, another prefers OpenAI's Codex. Pricing changes, and you switch without rebuilding. Local models get better, and you start running some workloads on your own hardware. The workspace is the constant. The tools are choices.
What gets harder The personal workspace is something one motivated person can stand up in a session. The team version requires thinking through the bigger picture of how work is shared and where data lives. What systems to connect and how deeply. How to scope permissions across different roles. What to automate and what stays manual. How to build skills that hold up for people other than the author. Where things get stored and how that should work, whether some data pushes back to a connected system, lives in a shared repository, or warrants a hosted database. And how to keep the whole thing improving over time. These are solvable. The teams getting the most from this are thinking deeply about each one: building reporting layers that let the AI query business data through a controlled interface instead of a raw database connection. Designing shared skill libraries where the best version of a workflow is tested, validated, and available to everyone rather than reinvented by each person. Building purpose-fit tools when a workspace alone is not enough, so the AI works through an interface shaped for the domain, not a generic chat window.
How we run ours RoleModel runs its own business on this pattern, not just in the codebases we develop for partners, but across business operations: sales, marketing, delivery, and strategy. Our CRM and project management tools are connected as data sources over MCP, alongside custom internal applications we have built for our own operations. On the skills side, the team uses shared skills for brand voice and document generation. A proposal gets drafted, reviewed against our voice standards, and laid out as a branded document, all within the workspace. A meeting gets prepped by pulling calendar context, partner history, and open action items in one step. These are not hypothetical. They run daily. We have built this for partners at varying levels of technical comfort. At its lightest, it runs as a "code" repository, no servers required, accessible through a desktop app or terminal. At the other end, portions can be hosted as applications the team reaches through their AI tools. The right shape depends on who is using it and what they need.
A
Appendix: Starter Prompts
You need an AI coding tool installed (Claude Code, Codex, or similar) and about 20-30 min.
Step 1: Create a project folder
Create a folder somewhere on your computer. This is your workspace. Name it whatever makes sense to you: personal-assistant, my-workspace, work. Open it in your AI tool.
Step 2: Paste the prompt and answer the questions
Copy the prompt on the following page and paste it into your AI tool. It will walk you through a conversational interview, one question at a time, and write the results into your workspace as it goes. Be specific. "I run marketing" is not enough. "I own demand generation, content strategy, and the website for a 40-person engineering services firm" gives the AI something real to work with.
Step 3: Use it
When setup finishes, run /plan-day to see your first daily brief. It will pull your real calendar, your real tasks, and prep you for your real meetings. From there, the workspace grows with you. When you want to add something, just say it in plain language: "I have a recurring 1-1 with Sarah on Tuesdays, start tracking an agenda for her." The AI will update the right files.
Day 1 vs. later
The interview only asks for what you need today. Things that grow into the system over time, recurring rhythms, agenda labels for specific people, quarterly priorities, are deferred. You do not need to have everything figured out before you start. Start with identity, your task tool, and your calendar. Everything else layers in as it becomes useful.
You are conducting a one-time setup interview to bootstrap my personal workspace. By the end of this session I should have a populated system/ directory with my context and working daily planning skills.
Style: One question at a time. Conversational, not a form. Wait for my answer, react, then ask the next. Push back when answers are vague. Write files as you go so progress is durable if we get interrupted. Be direct, no padding.
Step 1 - Identity: Ask my name, role, organization, working hours, and the 3-7 areas I am responsible for. Push for concrete answers. Write system/user_profile.md.
Step 2 - Personality (offer to skip): Ask if I have done DiSC, Working Genius, Enneagram, or similar. If yes, capture in system/user_personality.md. If not, skip.
Step 3 - Task management: Ask what tool I use to track commitments (Todoist, Linear, Things, Notion, a markdown file, etc.). Ask how it is organized. Check if it is connected as an MCP server. Write the result to system/user_profile.md.
Step 4 - Calendar: Ask for my work calendar ID (usually my work email) and personal calendar ID (or none). Check if Google Calendar MCP is connected. Write memory/reference_calendars.md.
Step 5 - Recurring rhythms (offer to skip): Ask about weekly, monthly, quarterly commitments. Capture prep templates if applicable. Write system/project_recurring.md.
Step 6 - Preferences: Ask what I want this system to do that nothing else does, and what I would NOT want it to do. Write system/project_assistant_system.md.
Step 7 - Create skills: Build three planning skills: - plan-day: Daily briefing pulling calendar, tasks due, and meeting prep - recap-day: End-of-day review closing finished work, capturing new commitments, looking ahead - plan-week: Weekly planning with calendar overview and priority setting Each skill should read from the system/ files you just populated.
Step 8 - Create AGENTS.md at the project root describing this workspace, and working in Claude a CLAUDE.md containing only @AGENTS.md.
Step 9 - Tell me setup is done and offer to run /plan-day right now.