What I’m thinking about: Claude Code sat through 18 sessions of a healthcare conference with me, feeding me questions as the speakers talked…and by day two, I wasn’t entirely sure which of us was the agent.
Every 10 minutes at MGMA, my MacBook told me what to ask, though I never once raised my hand (more on that below).
Last week I shared what medical practice leaders said about AI when nobody was pitching them, and promised the other half…how Claude Code and I reworked the conference plan at the last minute, and the superior (IMO) conference session setup we ran on today’s technology.
Quick recap: MGMA’s no-solicitation reminder landed moments before I planned to send 14 outside reads to target organizations (phew), so I went purely as a participant.
From pitching to learning, at the last minute
Once that email landed, Claude and I rebuilt the plan around learning and being a value-add to the host, MGMA. The printed reads and handwritten notes came off the list, and a learning agenda took their place, with ~18 sessions picked from hundreds over the 2.5 days I’d attend, chosen for where I’d learn the most, where I might eventually contribute, and who would be worth knowing.
Every session, from the 7 AM starts (healthcare runs on early risers) through ~4 PM, went onto my calendar with its room number, pre-seeded questions and Notion notes on the speakers.
The fire hose
The first session, on a Sunday afternoon, was two hours on revenue cycle management (how a practice bills, collects and fights denials), and it was overwhelming.
I hold an M.D., but I’ve run a real estate business full-time for the past six years and was never on the admin side of medicine, so I needed to get up to speed fast…I didn’t even know what people were talking about.
One typed sentence, mid-session
Notion’s meeting notes were recording the session live on my MacBook, with Claude Code running alongside (on Opus 5.5, released five days before the conference). Dictating via Wispr Flow (my usual process) wasn’t an option in a quiet room, so I typed my stream of consciousness, trying to live-adjust to my circumstances and what I thought would be most helpful.
(This is probably the #1 way I make so much progress with AI. I approach it with an open mind, and I don’t know if something is even possible until I get the wheels turning, as I’ve learned to stop making assumptions about its limits.)
I typed this into my Claude MGMA notes session:
“ok if you have live transcript, bring up q’s that are relevant to me as they come up. i’ll note questions if i want to, otherwise, bring up q’s every 10min or so, based on updates of what’s answered vs new directions”
Claude hooked into the live Notion transcript (through an MCP connection, the plug-in that lets Claude read and write inside another app) and put the question feed on a 10-minute loop. Each round read the newest stretch of the talk, marked which of my earlier questions got answered, and suggested new ones only when the speaker took a new direction.
(If you’d rather watch it run, the full screen-share is on YouTube, chapters included.)
In the same thread, at my direction, Claude defined the acronyms as they came up (111 by the end of the conference…healthcare takes the cake for the most jargon-heavy industry IMO, other than probably crypto).
As another example, when a government affairs town hall on Tuesday caught my attention, I typed “add the Tuesday town hall to my calendar,” and it appeared, room number and prep notes included (through Composio, which has a free plan and lets Claude manage more than one Google account).
That one Claude thread ran the entire conference and got extraordinarily long, and in my experience the newer models hold up FAR better on threads that long than they used to, as in they hallucinate much less and retain ‘focus’ better (though I’m still split on fresh versus long, and usually fork one thread per job as best practice ATM).
Worth bearing in mind if you try running simple prompts like mine above and don’t get nearly the same quality of results: my Claude Code setup has an extremely nuanced ruleset, backed by MCP connections and informed by thousands of words of constantly updated business context/memory, plus skills (saved playbooks) auto-called from a GitHub repo and hooks (automatic triggers) for Claude Code to cross-reference other threads.
In short, that always-evolving backend is why one typed sentence could run a whole conference.
Why I never raised my hand
Within that first session I realized I was paying a steep ignorance tax, and learning from what other people were struggling with would teach me far more than grooming myself as a contributor. As Stephen Covey put it, “Seek first to understand, then to be understood.” The rooms had plenty of questions without mine.
The question bank still paid off, sideways. Some questions entered our ‘seed bank’ for articles and FAQ answers on our site, built for AEO (answer engine optimization, getting cited when someone asks ChatGPT or Claude a question), and others pointed straight at AI business opportunities within healthcare, ranked by feasibility and need. By Tuesday the seed bank held 157 content ideas, with 10 opportunities ranked in Notion…and I didn’t have to carry any of it in my head.
What Claude couldn’t see
When a presenter asked for a show of hands and about a quarter of the room went up, that data point only existed because I saw it with my own organic eyeballs and typed it in…Claude and I have a partnership (though I never lose sight that I’m the only one who bears any risk).
Over the conference, Claude built its own routine. When a new recording started, it matched it to the session on my calendar, titled it in Notion for easy reference, and when the day’s sessions ended, it paused until the next day’s 7 AM start rather than burning tokens (the units AI usage is billed in) on dead space.
Not everything held up. Live transcription rapidly drained my MacBook battery in a convention center with very few outlets, so I pivoted to recording on my phone, which doesn’t transcribe until you hit stop (Claude noticed and paused the 10-minute question routine on its own). One recording also cut off after 15 to 20 minutes, and a couple of background agents stalled and had to be rerun (though these were one-time issues, each diagnosed quickly, with fixes now in place).
Claude read all of the presentation decks in parallel, using sub-agents that each handled a few, and merged them into the session notes (no more photographing slides…almost nobody reviews those anyway). After each session, a wrap-up researched the questions that hadn’t been answered and checked claims (within both the transcript and the slides) against 2025-2026 sources.
For example, Claude caught an attendee’s confident claim that Anthropic won’t sign a BAA (the business associate agreement HIPAA requires before a vendor touches patient data) unless you pay millions of dollars. This produced anxiety in me, as it could potentially throw a large wrench in my plans for weaving Claude setups into healthcare practices.
Once I investigated the claim, I found Anthropic does sign BAAs at a much lower cost, for HIPAA-configured Claude Enterprise, parts of its API and Claude Code, but connectors and MCP servers aren’t covered when data flows to a third party, a key caveat for heavily integrated setups like mine.
Fewer sessions, more meetings
Over 2.5 days, the returns diminished. I attended too many sessions with too much content overlap, and the best meat of the conference came from one-on-one conversations…hard to line up when you walk in at the last minute, knowing nobody at the conference.
The next conference faces a higher bar: I’d book meetings weeks in advance (or attempt to go with a well-connected partner/colleague who can facilitate intros) and cut the number of sessions I attend.
Still, that Claude Code thread made sure every MGMA follow-up is logged with a date, and the 14 outside reads I had already prepared go out next week as ordinary outreach, no longer tied to MGMA.
It reminds me of the elementary school lesson on how the Plains tribes used nearly every part of the buffalo they hunted…that’s how I approach conferences and AI output. Nothing goes to waste.
Machine head, human body
Riding escalators between sessions, with Claude reminding me I had 10 minutes to reach the next room, was the most surreal experience…I almost felt like an automaton, acting on Claude’s behalf.
The tech-forward crowd has a name for that feeling. In automation theory, a centaur is a person assisted by a machine, and Cory Doctorow describes the reverse centaur as a person serving as a “squishy meat appendage for an uncaring machine.”
That threw me at first, but I was still the centaur. I was in those rooms because of where I want this new business to go, and while Claude ran the tactical loop, the glossary, the calendar and its own overnight routine, it still couldn’t see a raised hand, walk a packed room, or decide which questions to keep to myself and which people deserved a follow-up. Ultimately, I decided what I wanted out of that conference, and the relationships that come from it stay mine.
None of what I built above came from a template, and I can’t see going to another conference without it, because I extracted every last nugget of value from each session while being MORE present in the room.
=====
If you run a ~$5M+ business, the work starts with deciding which jobs a system like this should carry and which stay with your people, then building it with verification where a human still owns the call. It’s how you become the operator who walks into every room already briefed. Everything about how that works lives at seriousai.partners, so reach out if this sparked an idea for your team.