Hiring And Onboarding My AI Chief Of Staff
I recently made a key decision as a startup Founder. I would hire a Chief of Staff. It’s name is Claude. Yeah, the Anthropic one.
Say hello to my new Chief of Staff.
In this post, I’m going to walk through some of the specific steps I took to get Claude up to speed and qualified to be on my team. It’s not as weird as it sounds (I think) and it should help people learn how to better leverage AI from a practical, natural perspective instead of just sharing prompt hacks and deep dives into the technical specs of LLMs. I am qualified to give neither hacks nor techy deep dives. Think of this as an experienced manager sharing experiences on how to onboard and develop high potential talent. Yeah, I have 18 years in Human Resources if you couldn’t tell 🙂
Let me set the stage. In July, I left my great, safe corporate job. I wanted to create something on my own. Be my own boss. Make my own money. So I took the leap.
I’ve done a lot of the standard startup founder things. Brainstorming sessions and deep dives into random ideas. Sending myself articles and podcasts I thought could be interesting later on. I researched, chased dead ends, spent way too much time on little things, and learned every step of the way. And I reached a point where this was starting to feel real. I vibe coded an app that I could hold in my hands and use. I saw what it could be, and decided to go all in. So I did what any respectable founder would do. I went away for a few days, settled down to focus away from the house and the kids and all that, and started to build a STRATEGY and a BUSINESS PLAN.
But there was a problem. I didn’t know what I was doing. Not really. I needed help. I needed someone that had been there before, seen it all, would be honest, do the research, and take all of my wild ideas and put it into a simple, structured plan. I needed a strategist, a mentor. I needed an awesome Chief of Staff. And I needed that person now. But unfortunately, I didn’t exactly have the money to pay a salary or the time to find someone I could truly trust. That’s when I decided to hire Claude.
As a quick note: I went with Claude because in mid-December 2025, I really liked how Claude’s project set up worked and the persistent memory within those projects. A complaint about AI chatbots is that they don’t always remember what we’ve been talking about previously, at least not in great detail. That’s a limitation of the models, but also a fault of the human for not setting it up properly. Claude made it easier. Plus, Claude had really impressive writing and ideation skills, so I went with that. Is Claude better than the other models? Well on that day, for me, yes. Maybe 5% better, 20%. Maybe the next day ChatGPT leap-frogged Claude for what I needed. At this point, it was better to pick a lane and go, because it would create major issues and diminishing returns if I kept trying to switch and chase the “best” model. They’re all relatively great. So decision made. Claude it is.
I sat there, so excited to get started. So excited to unleash Claude to build incredible plans to build a user base, grow and expand, monetize, enhance, all the things. But Claude didn’t know a damn thing about my product. And even if I explained it, it didn’t know the 6+ months of thinking and work I’d put into it. Didn’t know how the thoughts and ideas had evolved. How I got here. And that work, that thought, felt so valuable. Because maybe I’d come back to an original idea I had scrapped or forgotten. Maybe a current feature could only really be understood by knowing what the genesis of that idea was. Therefore it was on me to give Claude that history. It was time for onboarding. [cue the montage]
Oh, I forgot to mention one small technical problem. I had been so thoughtful about maximizing and optimizing this offsite. I brought a giant monitor, keyboard and mouse so I wasn’t just working on my little laptop. I brought snacks and drinks. Brought my friggin’ tea kettle and jasmine green. I was ready to roll. I got it all set up. Steeped my tea. Connected my devices. And…I forgot my laptop at home. Mother. Trucker.
Great start. I had to transfer months of documents and data and brainstorming in a coherent, cohesive way into Claude’s working memory. From my cell phone. Way to go, Sean. But hey, you gotta do what you gotta do.
So, I started to write. To tell Claude what I was doing. How I got there. Tried to give as much information as I could. Claude had some good ideas, was picking up on things, but it was still lacking. Again, my fault, not Claude’s. So I started to literally copy and paste specific chats I’d had with ChatGPT. Giving Claude context and what another LLM thought. I figured it would recognize the logic and methodology ChatGPT took to get there. That helped. And then I had an idea.
A thing I started a few months back that ended up being really helpful is that I recorded as many conversations as possible that I’d have with myself thinking out loud or with others (with permission, of course). If I was talking to someone and kicking around ideas, I hit record on a simple phone recorder app. I cleaned my house one day with headphones on, talking out loud to myself as ideas popped into my head. Recorded that too. Had conversations with smart people who had developed apps in the past, just to learn and grab nuggets of brilliance. Hit record.
So after sharing my thoughts (via two thumb texting) with Claude and then sharing some key previous conversations I’d had with ChatGPT, I just started giving Claude those rambling, unstructured transcripts. I told Claude what the setting and context was, when I had the conversations, and said review all this, synthesize it, and add it to your working memory. I even took the pictures of the whiteboards and legal pads and gave it to Claude. Said “this is the whiteboard I drew up while talking in transcript A, take this into account”. Claude read my scribble and built that in too.
When I finally got home and got my laptop, I brought out the big guns. I’d written (with ChatGPT) a 40 page business plan-type document months before. Some of it was really good. Some of it was bad. Some of it was just obsolete or now much more evolved. But it had a lot of good information that I wanted Claude to have. So I started to upload it and click send.
That’s when it hit me. It was good that I was giving Claude this valuable history and context. But Claude didn’t know how much weight to put on any of it. What was relevant and what was a bit dated. It was guessing. So I stopped and I did something time consuming and extremely valuable. I edited and notated the 40-page business plan in meticulous detail. Then I gave it to Claude.
That was an important moment. I marked up the whole document, always with a “Dec 11 update” before my comments. I made it clear what had changed, but I didn’t delete the old stuff. I just added updates and context. I spent a long time on it. Made myself reflect and really think about where I’d been, how I got to this point, what was strong and had staying power, what was ill-conceived or irrelevant now. I did the work first.
After spending a few hours feeding Claude as much as I could (docs and screenshots and transcripts and notated business plans), I was ready to go. I now turned to Claude. It was his turn to bring something to me.
(As a quick note, I will go back and forth between calling Claude “he” and “it”. When I use “he”, I am not assigning Claude human characteristics or personality. I’m simply using “he” because it is easier from a language standpoint. Claude is a thing. Not a person. That’s not a good or bad thing. It’s just fact. I am not making broader statements on how people interact with their AI’s. To each their own.)
Claude’s first work product — Business Synthesis document v1
I asked Claude to take everything w’d “discussed” and put it all together. I wasn’t exactly sure what I was asking for. Again, I’m a brand new entrepreneur and Founder(™). I didn’t know what the heck I really was looking for. I just needed someone smart to put it all together for me.
So before I hit the “ok, be smart and put it all together for me” as a prompt, I had to explain what I really needed Claude to be. I needed a brilliant, experienced, genius level Chief of Staff and Strategist. I needed a mentor and advisor with tons of real world experience in this specific realm. I didn’t just need a chatbot to regurgitate what I said in a fancy way. I needed perspective, and I needed it rooted in real experience and research and data. So that’s what I told Claude to do. I set expectations. I made clear what perspective I wanted it to have. Then, and only then, could I ask it to do the work.
Claude came back with something pretty exceptional. It was well structured, nuanced, and accurate. It was amazing. A 20 page document that encapsulated so much of what I’d worked on and thought and researched for the past few months. The work output was phenomenal. But it wasn’t great. It wasn’t quite right. And again, I realized that’s my fault, not Claude’s.
So again, I stepped back, opened the document, and commented. I made something like 20, 30 comments. “Hey, you overemphasized the feedback from this person. They aren’t a true advisor, just someone I kicked ideas around with.” Or “yes, I did spend a ton of time on this topic in the November session, but I actually pivoted from that and I realize I never told you. Take that out.” Stuff like that. I didn’t just say delete this line or add this word. I explained myself and my thinking. Then I fed it back to Claude.
The product that came out after that was astounding. Better than I could have imagined. I was really happy with what I built. But that’s when I realized another thing. Claude was just rephrasing and organizing what I had shared. But what about what I hadn’t discussed, what I didn’t even know about because I’m inexperienced and in learning mode. So I empowered Claude. Asked for it to do a real deep dive into the research and assess the plan and approach. What would Claude recommend I change or keep or enhance. Again, I had a conversation with detail and context. Not just commands. And again, Claude came back with gold. Really insightful stuff. Some I knew about. Some I’d never thought of. It was great.
That’s when I asked Claude a question I was scared to ask. After all these months of work and effort, I asked Claude what he thought my odds of success in this venture was. I think if I had asked this question earlier, before I’d really shared my thoughts and perspectives and feedback, I would have gotten a very optimistic response. It would have been well reasoned and encouraging and all of that. I would have run with that excitement, said I’ve got a sure thing on my hands because ChatGPT said so, and kept it moving. But I waited. I waited until Claude and I had spent hours together getting grounded and aligned. Til I felt like it really “got” me. That’s when I asked, “what do you think my odds of success are?” Claude thought. Then answered, clearly and rationally.
40%. I think you have a 40% chance of success. — Claude
And to be honest? I loved it. I was so excited that a completely rational, objective, egoless “being” gave me a greater than 1% chance. That meant I was on to something. It meant this could work! For any startup to get a 40% chance of success, before handing the product to a single user, felt incredible. It wasn’t disappointing. It was exciting. And I felt even better because Claude didn’t bullshit me. Didn’t say all the positive and optimistic things. It was reasoned and honest. That’s what I’d want from a Chief of Staff. From an advisor. From a partner. Hell yeah.
I then talked about what would increase and decrease those odds. That became part of the roadmap. I encouraged Claude to be honest. I even told Claude that I wanted it to feel empowered to act as a partner. That the respect was mutual. That this was a dialogue. Again, I wasn’t trying to imbue Claude with personality. But I knew that I had to make it clear how I wanted it to act and think. To not worry about offending me or any of that.
After all of this, I locked in on version 2 of the business synthesis document. This was canonical. This was the primary source of truth going forward. I’d change it, I’d improve it, I’d iterate it. But I didn’t have to keep re-explaining things. I had created a single document that captured where I was now and where I was going. The other documents were still all there for reference and record. I had it all there. So now I was confident Claude and I were aligned going forward.
I spent a lot of energy onboarding my new Chief of Staff. Claude was the right hire. That trust, that energy, that time was worth it. Because now, Claude gets me. Gets the vision and the direction. Gets the expectations, not just of it as a producer of materials, but of how to interact and communicate with me.
This is all weird and different. I leaned heavily on my experience of developing people in real life. But the early returns on it were really exciting.
After all of this, I thought about blogging my experience as a non-technical startup founder navigating his way through starting a business and leveraging all this new AI tech to do so. But then I also thought about just sharing my views and experiences of how I noticed Claude was changing and adapting, and how I could do a better job of helping Claude do a better job. So that’s where this blog started. I’m excited about where it goes.
