Hot Take: GrokBot can't do anything new that Claude can't do. Here's what the @bot team DID do that's making it so popular (IMO): They demystified "building a team of agents." You'll see people all over the internet talking about how they "build agents with Claude." That could mean that they: -Built a skill -Created a folder with a Claude.MD file and some extra context -Built a Claude project -Coded their own harness -They're using managed agents thru API -Connected Claude to some MCPs -Set Claude to do something on a timer -They're using subagents -They could even just be talking about having multiple chat threads going with Claude -Or any combination of the above Here's the problem: They're all correct. No one knows what the heck anyone is saying when they tell you they "built an agent with Claude", or that you should build one. Even @AnthropicAI is ambiguous about this. I routinely get marketing emails from Claude Code marketing team talking about how they're team is using teams of agents they've built, and it's extremely unclear what they've actually done. Now they've launched @claudeai Duet. But they're still just giving you one agent. With GrokBot, it's a simple and easy to understand what it means to have a team of agents. An agent gets it's own chat. It gets it's own skills & routines. It gets its own context. It doesn't have ALL of your context, but it can get additional context from other agents when needed, and all your agents can chat with each other as needed to get stuff done. Clean, simple, unambiguous. If someone tells you they built a team of agents on GrokBot, you know what they did, and if you open up GrokBot, it's very easy to see how you can do it too. End rant.

This week I used @bot to launch a free nationwide database of publicly announced CRE lease & sale transactions. It's entirely automated through GrokBot. You can check it out at dealrevue.com. 500+ website hits later, here's how it's going for anyone interested in using GrokBot to automate something useful (read to the end for the key takeaway on AI generated content.) Quick 🧵 How it works 👇 (P.S. You can also sign up to receive a weekly rundown of announced deals in your market direct to your inbox.)

OpenClaw comes out: You can build a team of agents! Claude: Here’s dispatch so you can run Claude from your phone. GrokBot: You can build a team of proactive agents that actually work now! Claude: Here’s one proactive agent @claudeai why the half measures?

🚨 1 slam-dunk use of Claude for CRE? Due diligence checklists that fill out themselves, and find conflicts across your deal docs. We built a Claude skill to tackle it. On the first run, it asks for YOUR checklist template. Excel, Word, PDF, or pasted in. Point it at the deal room folder and it reads every file. You get one workbook back. Every item in the checklist marked Received, Partial or Outstanding, and a conflicts tab outlining any disagreements across deal docs. The PSA lists three parcels but the title commitment only covers two? Flagged. Deal rooms grow overtime, so it offers to create a scheduled task calling the skill like clockwork. It checks the folder for new docs, updates the list, adds or resolves conflicts, and emails you with an update. You keep the judgment calls. It takes care of the hours of assembly. Grab the Due Diligence Checklist skill & our full skill library in CRE AI Studio: creaistudio.com

Weekend GrokBot project: A free nationwide database of publicly announced CRE lease & sale transactions 1 bot finds new transactions 1 pushes the new pages live 1 updates weekly summaries for each metro 1 analyzes the data & makes TikToks to get the word out with Claude/Higgsfield/HeyGen, & pulls engagement metrics to a data pipeline so they improve overtime. Haven’t touched my computer once in the process. Check it out here: dealrevue.com

Over the past 4 weeks I’ve delivered 9 in-person AI workshops & presentations at commercial real estate conferences across the country. Here’s a few takeaways: -Most weren’t sure what MCP is -Most had heard of skills but never made one -Everyone had heard the term AI agent but were unclear on what makes something an agent, or the fact that they already probably already had access to one -Most had no idea what deep research/research mode was or the fact that they’ve had access to it for almost 2 years Don’t let social media fool you, It’s early.

One full week in Denver, CO. What’s the number 1 thing I should do/see/eat/drink?

ChatGPT Astra is the first time I've seen a model build a commercial market survey in dang-near one shot. -Gave it for random properties to find intel on (though it'd be more efficient to just hand it the brochures) -API to build the map -Bullet points on functionality -Brand dot MD file Produced an on-brand, interactive survey in about 20 minutes. Would it need some tweaks? Probably, but way less back and forth needed than I've ever seen before. This is my typical approach with model releases. What can you do now that previous models have failed at? Or, what can you do faster and with less friction than ever before? Full tutorial, prompt & API config available in CRE AI Studio PS: Why does no one talk about the convenience of the ChatGPT Sites plugin?

2 days in Chicago coming up. What’s the one thing I shouldn’t leave with it doing? (Or eating)

When someone tells you they "built an AI agent", be skeptical. I had GrokBot pull the top 100 Youtube videos on "building AI agents with Claude". Here's what they were actually building: Most (31%) were just building skills (admittedly, I love skills) 20%+ were just talking about using Claude Cowork on desktop/an overview of how it works. (Claude Cowork is an agent.. but did they really build an agent?) 12% were talking about connecting Claude to no-code building tools like Base44 and other external platforms. Only 31% were talking about creating subagents/agent teams or managed agents (tho Grok did bucket some videos about creating agent MD files into a broader bucket of creating MD files.) Don't obsess over "building agents." Someone telling you to do that could mean many different things. Build the systems that actually support your work.

A few years ago @hubermanlab said reading fiction is good for your brain. I never enjoyed reading but gave it a shot. The more I read the more I enjoyed it. This year for the first time I’m at over 1 book a month. Mental clarity feels like an all time high. Business books are great. But if you’d like better sleep and mental clarity, I highly recommend picking up some fiction.

2 weeks of putting GrokBot to the test. Here's what I wish I knew day 1: 1/ The hype is real around it being easy to setup. Making an agent, connecting tools, it's just easy. All happens in the chat. Easy to use from your phone. This is the big draw, it's accessible. 2/ It's not a magic button. Easy to set up & easy to automate real work are 2 different things. I tried my best to treat it like an AI newbie. Reality is, unless it's doing something pretty simple, most of what's required to get a quality output from your go-to AI tool still applies. Give examples, evaluate, iterate until it gets something right. Then automate. 3/ For the reason above, don't immediately launch 10-15 different agents. You'll end up with a mess of automated work, none of it going well. 4/ Start with a Chief of Staff agent. Tell it about your business, connect it to your tools, have it audit, analyze, and provide suggestions on ways it can help. @gregisenberg did a great podcast on this approach. 5/ Pick the ones you ACTUALLY think will help your business. This is the trap, GrokBot can do a lot. It's so fun and cool to see it happen that it's addicting. You can easily wipe out your weekly credits setting up agents and routines that have no impact on your business. You'll be upgrading to the $200/mo plan in no time if you take all of it's suggestions. If you don't care good on ya bit if you won't want to blow $200/mo on a whim, pick your battles. 6/ Make the picks and focus on ONE of them with the Chief of Staff. Once it's done something that's actually useful, it can pass it off to a new agent. Chief of staff gets it done right the first time then passes off. Again, setup is easy but this part could take longer than you think and that's fine. 7/ Once the task is passed off to a new agent, it can then set up a routine so it pulls the info from your connected tools and does the work weekly/daily/ whatever cadence you need. 8/ "When do I need a separate agent?" You could probably ask 10 people and get 10 different answers but from my experience it's when enough tasks center around a specific domain of your business, and an agent should have it's own memory/context that isn't needed in other domains of your business. The trade off between specialized and generalized. 9/ "SHOULD I SWITCH?!" Don't focus on switching from your current platform. Test it. See if GrokBot can do something beneficial for you or your business that you've struggled to get your AI platform or choice to do, or if it can be done better with a team of agents focused on specific roles/ tasks than with one mega agent with endless context. You might switch, you might want both. Only one way to find out.

I built a retailer map maker skill with Claude Cowork. Subject property. Branded logos of every nearby tenant. Clean satellite view. First map looked great. Except some of the logos were stacked on top of each other. So instead of asking Claude to fix that one map, I asked it to fix the skill: "Bake in an edit mode. I want to drag logos around. Resize them individually. Highlight the ones that matter, change colors, and export to PDF." It did it. Now every map the skill generates comes with: Click-to-edit mode Draggable, individually sized logos Bulk resize to clean up clutter in a single action Best of all, upload your brochure and it builds the entire map around your brand colors. No design software. No re-prompting. No back-and-forth for every map. Here's where we are with AI now: You aren't just generating outputs. You're upgrading the tool that generates them. In plain English. In the same conversation. Full downloadable Claude skill at creaistsudio.com

I get more requests to build AI skills for CRE teams than I have time to execute on. Here's the framework I use with them so you can do it yourself. It's not rocket science. It's time invested once so your whole team saves time later. Building a skill for yourself is simple. Have AI do something for you. Iterate until it gets it right, then ask it to bundle the skill. Boom, you can now do that thing on repeat just by using the skill. If you hit roadblocks using it, you can always update the skill as you go based on nuances/ outliers. A skill a team can use is a little different. A - You probably don't want everyone across your team riffing on skills in real time, you end up with 10 different versions of the skill all working differently. B - If the skill doesn't seem to work for them or takes more time to use than just doing the task themselves, they won't use it. And for good reason. Here's how to build a skill you can roll out. Example: BOV generator skill. Step 1: Grab 3 finished BOVs you/ your team already produced. BOV A, BOV B, BOV C. The fresher the better so you have the inputs handy. Step 2: Dig up the actual inputs you/ your team used to actually create these BOV's. Comps, property intel, analysis, etc. Note: Just because AI can perform it's own analysis doesn't mean you should skip your own analysis. Step 3: Give AI BOV A's inputs, your instructions on how to take the inputs and extrapolate a BOV, and BOV C as the template/example of what the finished looks like. Do NOT give it BOV A. That's the answer key. You grade against it. You don't hand it over. Step 4: Before AI starts, tell it to ask you every question it needs answered in order to execute the task. This helps identify gaps in your inputs/ places in your instructions it would need to guess. AI's getting pretty good at doing this naturally, but just asking it to do this can save you time by giving it the intel up front. Step 5: AI produces its own version of BOV A (BOV A2). Compare BOV A (your work) against BOV A2 (AI's work) line by line. This is the evaluation, when I spend time building skills with teams, this is where most of it goes. The more thorough the evaluation the more likely it'll nail it on the next try. Then give AI the feedback in one prompt, or on the marked up version of the deliverable (works well with spreadsheets/ reports, you can add a column for line by line feedback): what it got right, what it got wrong or missed, and any EXTRA information it provided in the end result that you just don't want (extra info is often more prevalent than incorrect/missing info). For the stuff it missed/ got wrong, ask it why. Often the reason is additional gaps in the inputs or instructions, or it made assumptions you don't want it making in the future. Step 6: Have it try again with the feedback in mind. Most of the time it nails the second pass. If it does, tell it to lock everything it learned into the skill. Step 7: Run it again. Give it the inputs for BOV B, this time use the skill. Compare AI's work against the real BOV B. Most of the time, if you've followed the steps, it nails this next run. But run step 7 as many times as needed against real inputs and grading against real finished examples until it's nailing it on repeat. The more nuanced/outliers there are in the process the more examples you'll need to run to make sure it's catching them. And if there's endless nuance/outliers, you may want to repeat the whole process for separate skills (IE BOVs for different types of properties). It's not rocket science, it's time. The time spent on the evaluation and feedback is the difference between a skill you need to fix as you go and one that works. Hand your team something that works.

The highest leverage tip I can give you for building Claude Skills: Don't build the skill up front. You'll end up making endless revisions and updates, reinstalling the skill over and over, and wasting a ton of time. Instead, just work with Claude to actually execute the process from start to finish, without a skill. Make sure it has access to the tools it needs to get the job done. Iterate on the final output until you have it in a final format that you can actually use. Then have Claude build the skill around your conversation, in light of the bottlenecks you've already discovered, the feedback you've already provided, the tools you thought were already connected but had to set up. This'll separate your skills that actually provide value from ones that just burn tokens.

Every acquisition starts with the same question: who owns this? I built a Claude CoWork skill that finds the answers for you in minutes👇 Here's how it works: 1. Paste an address (or send it from your phone via Claude Dispatch). 2. Claude CoWork fires up Chrome and walks the public records like you would: county website, city assessor, state business filings. It's built to find the sources in your market on the first run and save the pattern for later. 3. If a site is blocked for some reason, it falls back to more general sources to find decision makers, corporate websites, BBB profiles, LinkedIn lookups, news, etc. You don't get stuck just because one source is down. 4. It hands you a structured ownership report: deed-holder, controlling entity, the actual human decision-maker you can reach out to. Tested across NYC, ME, WI, MA and CT so far. Every time it runs it gets smarter and takes less time to fire, as it tracks the nuances of finding the intel in your market. Full downloadable skill at CRE AI Studio. Grab your 7 day free trial at creaistudio.com

This weekend I built a CRE tool I've been thinking about for a while: CRE demand across asset classes, visualized nationwide based on real Google search data. I'm looking for a handful of CRE pros to test drive it before it goes live. Here's how CRE Demand Seeker works: Toggle between volume, year-over-year change, or 3-month momentum. Filter by asset class. For sale and for lease search terms are separated. Consumer rental demand is its own layer. Click any state, then any metro market, to see how search demand for each asset class is moving in your market. Drill into the individual search terms feeding the data ("office for lease" vs "office space for rent" etc.) Here's 3 things that stood out to me from CRE Demand Seeker's April data synthesis: Office for SALE searches are up +49% YoY nationally. Office leasing search demand is up, but moreso in specific markets, it's down in many states. The buyers are showing interest in a much bigger way. Self Storage: Sale interest is overall up YoY (+5%). But consumer demand for RENTING self storage is down 22% nationwide. Buyer interest still has momentum while rental searches are disappearing. The state with the fastest growing demand based on CRE searches? Kansas. Much smaller volume than the heavy hitters, but interest in nearly every asset class is growing here. If you'd like to take CRE Demand Seeker for a test drive, let me know and I'll send it over.