Every shiny new technology shows up the same way: as an answer wandering around looking for a question to attach itself to.

That's roughly where coworking is with AI right now. The conversations tend to start at the end. How do we use AI? Where do we stick an agent? What can we automate? All very sensible. Possibly all the wrong place to start.

Because no operator I've ever met has an "AI problem." What they often have is a pile of enquiries nobody's answered, follow-ups that go out a day too late, the same support question asked fifteen different ways, onboarding that depends entirely on who's working that morning, and one brilliant community manager who keeps the whole place running out of her own head (and who is, naturally, about to go on holiday).

AI is only worth anything if it helps with that. The rest is just a nice demo.

Which is more or less what Uri Levine is getting at with his line: fall in love with the problem, not the solution. The solution is the fun part: the demo, the launch, the screenshot you fire off to the team at 11pm. The problem is the boring part. The problem is also where all the money is.

It's usually where the wheels come off, too. The first demo dazzles. Then the answers go a bit generic, the tone's slightly off, and the whole thing falls over the moment something specific happens. Cue the verdict: AI isn't ready for us yet.

Maybe. But more often it's the other way round. We're not ready for it.

Not because the team's behind, or the tech's too clever. It's that an agent needs two fairly dull things first, and almost nobody bothers to set them up.

It needs context: actual knowledge of your spaces, your members, your prices, your policies, how you talk, what's happening today. The stuff your good people simply know.

And it needs skills: the how-we-actually-do-things. The way an exception gets handled, when to escalate, what a good answer looks like.

Without context, it guesses. Without skills, it improvises. Give it both, and it stops being a chatbot with delusions of grandeur and starts being genuinely useful.

So before asking where to "put AI" in your space, here's the better question, and the one behind this whole issue:

What are you already doing every single day that would be faster, more consistent and a good deal less painful if the agent actually understood your business?

Let's go find it.

Large coworking operators and boutique spaces alike now have access to the same large language models (Claude, ChatGPT, Gemini). They are broadly available, and the performance gap between models keeps narrowing.

Two coworking spaces can run the same AI model and get completely different results. Not because one prompt was more detailed, but because one agent was built on deeply defined business knowledge and the other was not.

So what does context actually mean?

In coworking, as in any industry, context is not a single document or markdown file. It's a structured set of foundational definitions that shape how a space operates. There are three core categories to start with, and every operator needs to define them before anything else:

Context lives in two distinct layers. The first is static: foundational definitions that don't change often (brand voice, policies, space inventory) typically stored in structured documents like .md files that agents can reference consistently. The second is dynamic: live operational data tied to the member lifecycle (who the member is, what plan they're on, their usage history, recent activity) which requires a reliable, real-time data source to be useful.

Neither layer is sufficient on its own. Together, they give an agent both the rules of the business and the data context of each specific situation.

Context can't be built without reliable and unified data about the entire member lifecycle.

Most operators work with fragmented data: websites, CRMs, billing systems, access control, bookings, community tools, and marketing automation all store information in separate silos. This fragmentation makes it difficult to answer basic questions about members, their journeys, or how spaces are actually used, and it's precisely this operational reality that prevents AI agents from doing relevant work. That's where Nexudus comes in.

Nexudus acts as a unified data platform for coworking and flex operators, a single source of truth that consolidates customer and space data across the entire lifecycle, from first touchpoint to daily usage and long-term retention. It integrates data from access control, payments, accounting, Wi-Fi, bookings, community activity, sensors, visitor management, and downstream tools like Slack, Zapier, and HubSpot.

The result is a 360-degree view of the member journey, connecting marketing touchpoints, physical space usage, transactions, and support interactions into a single picture. And that's precisely what creates the conditions for shared context where AI agents can generate work that's actually relevant.

In a world where models are becoming commodities, the real differentiator is structured context: a structured definition of how the space operates and live data access across the member lifecycle. Nexudus provides the foundation to build both.

With the data foundation in place, the next question is how to turn that context into action. That's where skills come in.

With the context layer in place, the next step is to codify the workflows that put that knowledge to work.

Every experienced team member develops a unique set of working patterns. A salesperson can assess a lead and instinctively know how to open the conversation. A community manager can scan a queue of support tickets and immediately identify which issue to prioritize and how to respond. These patterns emerge through repetition, accumulated judgment, and a deep understanding of how the business operates. Yet they are rarely documented.

Most operational knowledge in coworking businesses lives in resolved support tickets, in email threads between team members, in the tribal knowledge passed informally from a senior community manager to someone new to the role. It exists, but it isn't structured, and that means it can't be transferred to an agent.

For most operators, the natural starting point for deploying AI is data access: connecting an agent to the CRM, the help desk, the booking system. Data access does matter. But it isn't enough on its own. An agent with two years of resolved support tickets doesn't automatically know how to classify a new one. It doesn't know which issues are urgent, what tone to use with an enterprise member versus a hot-desker, or when to resolve and when to escalate. Having the data is not the same as knowing what to do with it.

That's where skills come in.

A skill is a structured document that teaches an agent how to handle a specific type of task.

A skill captures a repeatable way of working: the steps someone follows, the judgment they apply, the decisions that have been made over time. It can include step-by-step instructions, success and failure examples, edge cases, hard constraints, escalation criteria, and reference outputs from the team's best work.

Skills make a workflow visible and repeatable. They are the mechanism by which institutional knowledge gets transferred from the people who hold it to an agent that can apply it consistently.

To understand how skills work in practice, it helps to look at what one actually contains.

In Claude, skills are built as a folder containing a SKILL.md file. That file holds everything the agent needs: metadata, instructions, examples, reference materials, and constraints. The structure matters because Claude uses it progressively, it reads the metadata first to determine whether the skill is relevant, then loads the detailed instructions to execute the task. In Coworkings AI #008, we dived into what a Claude Skill is.

A well-built skill for coworking operations typically contains:

The skill file can be supplemented with additional reference materials (taxonomy files, example outputs, response templates) stored in the same folder and referenced from the SKILL.md. This allows the agent to access supplemental information only when the task requires it, without loading everything at once.

Taken as a whole, the skill library is an operating manual for the business.

Context and data access tell the agent how the business operates and how members behave within the space. Skills define how specific tasks get done. Together, they are what turn an AI agent from a generic tool into something that can actually run operations.

Before writing a single skill, this is the process we recommend operators follow:

Step 1: Map the repeated work. Identify the tasks your team executes most often in a week, high frequency, predictable pattern, limited discretionary judgment per case. Start by describing exactly how a human does it today. If you can't describe that, you can't write a skill for it. For each task, document what happens from input to output: every step, every decision, every exception. What input arrives? What gets evaluated first? What decisions are made? What output is produced? What happens when something goes wrong?

Step 2: Evaluate the team-produced outcome. This is your baseline, the quality level the agent must match or exceed. Collect 75–150 real examples of how your team currently handles that task: resolved tickets, sent following-up emails, etc. This serves two purposes: it establishes what doing the job well actually looks like, and it becomes the reference material you use to train the agent.

Step 3: Define success and failure explicitly. Establish your acceptance criteria: tone, response relevance, information accuracy, policy compliance. Assign a simple rating (bad / average / good) to each criterion, and document what each rating looks like in practice. Without this step, improvement becomes impossible to measure.

Step 4: Write the skill and validate with a small sample. Build the SKILL.md with role definition, step-by-step workflow, reference outputs, escalation logic, and constraints. Then run the first 20–30 real cases through manually. Score each output against your acceptance criteria, and note what fails and why. Do not automate until quality is consistent across the full sample.

Step 5: Iterate. Incorrect outputs become the input for the next iteration. A well-documented error is the most valuable instruction you can give the agent, so capture it and fold it back into the skill definition.

In Coworkings AI #012 we walked through this entire process in detail using a real customer support example. If you haven't read it yet, it's worth the time, the article is practical and full of concrete examples.

Once the first skill is deployed and has passed manual supervision on a small test batch of real cases, monitor how the agent continues to perform and then pick the next workflow in your operations to build a second skill. The skill library grows from there.

Two skills we've covered in previous Coworkings AI editions make this concrete:

In the coming editions we'll keep sharing new skills and AI workflows so operators can use and deploy them easily.

The first version of any skill is necessarily imperfect. It was built on the team's best description of a process, on reference outputs that represent what good looks like today, and on assumptions about the cases the agent will encounter. Some edge cases won't have been defined in the skill.

The way to improve a skill is not to spend more time perfecting the first version. It's to deploy it, read the outputs carefully, and let the errors tell you what the next version needs.

Each failure is an instruction. When an agent misclassifies a support ticket, that failure documents a gap in the classification or escalation logic. When it recommends the wrong plan to a tour booking lead, it reveals a missing input variable. When the tone of a member response lands wrong, it points to a specific gap in the brand voice guidelines.

The iteration cycle has four phases: define the initial context and knowledge base; write the first version of the skill; analyze outputs to identify what needs to change; update the skill incorporating what the errors revealed.

This cycle is what makes the context layer and skill library valuable.

AI adoption in coworking and flex spaces is still slow, not because the technology isn't ready, but because the industry needs more practical guidance on how to implement it properly. A solid, trustworthy playbook is what builds operator confidence, and operator confidence is what drives adoption.

Defining context and skills is the first step toward an effective AI strategy:

Once the context is in place, skills are what connect that knowledge base to real automation. Start with a single skill, the most repeated task in your operations, and build from there.

At Nexudus, we are building this inside our AI agent platform to make it easier for operators to get value from AI faster.

We'd love to hear how you're approaching AI in your space, let us know what you think.

That's it for today. See you in two weeks! 🙂