The first time I set up Connect, I tried to move my whole business into it at once. Customers, content, operations, standards — everything. Most of it sat empty for weeks while I figured out which parts actually mattered.

Here's what I tell people in onboarding now: pick one workflow. Not the whole company. One workflow.

A Waxell Connect workspace is a folder an AI agent reads on its own — the files inside it become the agent's working context before you type a word. The fastest way to understand what that means in practice is to plug one real workflow into it and use it. Content creation. Hiring. A newsletter. Customer onboarding. Marketing campaigns. Whatever you actually do every week. Once one workflow runs in Connect, you see the pattern for everything else.

Why one workflow, not the whole business

The temptation is to build the system first and use it second. I did that, and it's the wrong order. You don't know which context files matter until you've tried to produce real output from them. You don't know what belongs in the playbook until the agent does the work wrong and you fix the instructions.

One workflow gives you a tight feedback loop. Build the workspace, do the work, see what's missing, fix it. By the end of the first week you'll know exactly what a good workspace looks like — and you'll have built one that actually runs, not five empty shells.

Choose the workflow

Pick the one you do most often that involves repeating context. The sign is that you keep re-explaining the same thing to your AI: here's my audience, here's my voice, here's the project, here are the constraints. If you're pasting the same brief into chats every session, that's the one.

For me it was content creation — the blog. For other people I've onboarded, it's been hiring pipelines, newsletter production, customer outreach sequences, or marketing campaigns. The specific workflow doesn't matter. What matters is that you pick something real, something you'll use this week, not something aspirational.

This is a good place to bring in your external agent. If you're using Cowork, you can talk through the choice with it — describe your week, what you spend time on, where you keep re-explaining context — and it can help you identify which workflow will get the most value from Connect first. The planning conversation is useful even before you build anything.

Make a workspace for it

One workspace, scoped to the workflow you chose. Not "Frances's stuff." Not "Marketing." Something specific: "blog," "hiring-q3," "newsletter," "customer-onboarding."

Your external agent can create the workspace for you. In Cowork, I say "make a workspace called blog in marketing" and it does. No clicking through a UI. If you're working through the API or your own tooling, same thing — one call and the workspace exists.

Get basic context files in there

This is the part that replaces the paste-the-note step. Think about what you copy into AI chats every time you work on this workflow. For the blog, mine were: brand voice guidelines, a content log tracking what's been published, and a list of internal links I can use. For a hiring workflow, it might be the job description, the interview rubric, and notes on what the team needs.

Start with two or three files. Plain markdown. They don't have to be polished — they have to be true. You can revise them once you see how the agent uses them.

Your external agent can write these too. I draft most of my context files in Cowork. I describe what the agent needs to know, and Cowork writes the file and uploads it to the workspace. The first version of my voice guide was a Cowork session where I said "here's how I write, turn this into a standards doc the agent can read." Took ten minutes.

If a piece of context changes often — what you're focused on this week, a customer's status, a running list of topics — make it a state object instead of a file. A state object is a live record the agent reads and updates as work happens. You can have your external agent create state objects too. In Cowork, I describe the data shape and it builds the object in the workspace.

Write a playbook

A playbook is a markdown file the agent reads when it enters the workspace. It tells the agent how to work: what the purpose of the workspace is, what process to follow, what standards to apply, where the relevant files live. A playbook differs from a prompt by location. A prompt lives in your head and you retype it. A playbook lives in the workspace and the agent finds it.

Your external agent can draft the playbook for you. This is one of the highest-value things Cowork does in my setup. I describe the workflow — what I'm trying to produce, what the constraints are, what the agent should avoid — and Cowork writes the playbook and uploads it. The first draft is usually 80% right. I fix the 20% after the first real run.

Keep it short to start. Four or five lines. Purpose, audience, process, standards. Fix it when the agent produces something wrong. That's how you learn what the playbook actually needs to say.

Start creating

This is the step people skip when they're in setup mode, and it's the most important one. Use the workspace to do actual work. Ask the agent to produce the thing your workflow produces — a draft, an outreach email, a candidate summary, a newsletter, whatever it is.

In my setup, I open a Cowork session, specify the workspace, and the agent enters it and reads everything before I type a word. No copying, no pasting, no re-explaining the business. The agent already knows the voice, the context, the constraints, because it read the files.

The first output won't be perfect. Good. That's how you find out what's missing from the context files and what's wrong in the playbook. Fix those, run it again. By the third or fourth cycle, the workspace is tuned to the actual work, not to what you imagined the work would need.

This is my specific setup — I use Cowork as my interface for Connect. Connect is also accessible via API and web UI, so if you've built your own agent tooling, you get the same behavior. The agent reads the workspace on entry regardless of how you reach it.

The pattern teaches itself

Here's what happened after my blog workspace was working: I looked at my customer outreach workflow and immediately knew what to do. Same shape. Workspace scoped to the workflow. Context files with the information I kept re-explaining. Playbook with the process. The second workspace took a fraction of the time because the first one taught me what mattered and what didn't.

That's the actual onboarding strategy. Not "set up everything." Set up one thing. Use it. Learn the pattern. Then apply it everywhere else.

The cost of not doing this is measurable. LLMs are stateless by design — no context carries from one session to the next, which is the structural reason your agent forgets when you close the tab (Atlan). And the tool-switching tax is real: one 2025 survey of 1,000 U.S. knowledge workers found the average worker loses 51 minutes a week to toggling between tools, over 44 hours a year (Lokalise). The paste-the-note loop sits right inside that number. One workspace kills it for one workflow. Then you do the next one.

If you want to try it, you can get access at waxell.ai/get-access.

FAQ

What's the best way to get started with Waxell Connect?

Pick one workflow you do every week — content creation, hiring, customer outreach, newsletter, marketing — and build a workspace for it. Add the context files you keep re-explaining to your AI, write a short playbook, and start doing real work in it. Don't try to set up your whole business at once. One working workspace teaches you the pattern for everything else, and you can have your external agent (like Cowork) help at every step: choosing the workflow, creating the workspace, writing the context files, and drafting the playbook.

Can Cowork or an external agent set up my workspace for me?

Yes, and this is the easiest way to do it. Your external agent can create the workspace, write the context files, draft the playbook, and create state objects — all from a conversation where you describe what you need. In my setup, I use Cowork to do all of this. I describe the workflow and what the agent needs to know, and Cowork builds the workspace, writes the files, and uploads them. The first version is usually close enough to start working with, and you refine it from there.

How do I get an AI agent to remember context between sessions?

You put the context somewhere the agent reads on its own, instead of pasting it into each chat. In Connect, that place is a workspace: files and state objects the agent consumes when it enters, before you give it an instruction. Language models are stateless — nothing carries from one session to the next by default — so persistence has to come from where the context lives, not from the model. Once the context is in the workspace, every session starts with the agent already knowing it.

What's the difference between a state object and a document in Connect?

A document is text you write and rarely change — brand voice guidelines, product details, process docs. A state object is a live, versioned record the agent reads and updates as work happens. Use a state object for anything that moves: a customer's lifecycle stage, this week's priorities, a running task list. The practical difference is who keeps it current. A document: you. A state object: the agent. Your external agent can create both — describe the data and it builds the object or writes the file.

Do I need Cowork to use Connect?

No. Cowork is how I work with Connect — I open a session, name a workspace, and the agent reads the files automatically. But Connect is also accessible through an API and a web UI. If you've built your own agent tooling or you prefer the browser, the workspace behaves the same way. Cowork makes the setup faster because it can create workspaces, write files, and draft playbooks for you in conversation, but it's not required.

How long does the first workspace take to set up?

With an external agent helping, about thirty minutes from start to first real output. Most of that is the conversation where you describe the workflow and what context matters. The agent handles the creation. Without an agent helping — doing it manually in the UI — budget an afternoon, mostly spent deciding what to include. Either way, resist the urge to make it complete. Start with two or three context files and a short playbook. The workspace gets better the first time you use it and see what's missing.

Sources

• Atlan. Why AI Agents Forget and How Persistent Memory Fixes It. https://atlan.com/know/why-ai-agents-forget/

• Lokalise. Too Many Tools, Too Little Time: How Context Switching Is Killing Team Flow. https://lokalise.com/blog/blog-tool-fatigue-productivity-report/