What learning AI at 51 actually looked like (I’d been using it since 48)

TL;DR: I started using AI at 48 because I was afraid it would take my job. For the better part of three years I was busy with it – ChatGPT, custom GPTs, a capture system, lots of motion, not much progress. What changed wasn’t a course or a better prompt. In March 2026 I switched to Claude, and in April – at 51 – to Claude Code, and within weeks I was building things I’d failed at for years, because the tool could finally work in my real files and remember my work between sessions. This is the dated, honest version of that story, flops included.

Where I started

I’m a scientist by training and a science writer by profession. My clients are regulated biotech and pharma companies, the kind of work where being wrong has consequences. When ChatGPT showed up, I looked at a machine that wrote fluent English about technical subjects and recognized the job description. Mine.

The fear was specific. My professional edge was doing the scientific writing in native English, and here was a machine handing something like that edge to anyone with a laptop. Maybe that fear was ridiculous. It got me moving anyway, which is more than most of my reasonable thoughts have done.

So I didn’t wait to be convinced. At 48 I decided I would not be the person who can’t work the new thing, and I started using AI – not daily, if I’m honest, but regularly, and on everything. Work drafts, hobby projects, nonsense questions, all of it partly one long experiment: what can this thing do, and what can’t it. That decision I got right. What I got wrong took me almost three years to see.

The treading-water years

For the better part of three years, my AI use was busy and went almost nowhere. I thought it was helping my productivity. Now I’m not so sure.

The effort was real. I built custom GPTs with my style guides loaded in – my ideals, my processes, the way I do SEO. The output never made me happy. It sounded reasonable, then needed so much manual work afterward that I sometimes wondered what the machine had contributed – and hallucination was a constant worry, because of what I do for a living.

The deeper problem took me longer to name. When I wrote about a topic I didn’t know deeply, ChatGPT would hand me pleasant prose, and I couldn’t connect my brain to it. I hadn’t researched the thing, hadn’t outlined it, wasn’t turning it over in my off-hours the way you do with a story you own. The draft was fine and it wasn’t mine, and making it mine cost about what starting from scratch would have. Nice sentences, added time, and a low-grade dissatisfaction I couldn’t put down.

And I want to be fair to those years, because they taught me the one thing I still care about most. ChatGPT kind of knew me. After enough conversations it carried something of my projects and my preferences – and it also kind of didn’t, which was its own strange experience: close enough to feel known, wrong often enough that I couldn’t relax into it. When I tried an AI without even that, I felt the loss immediately – my actual complaint at the time was “Notion AI doesn’t know me.” An AI that knows your work is a different tool from an AI that doesn’t. I’d found the right thing to want. I didn’t yet own a way to keep it.

One habit from those years turned out to matter more than every tool I tried. About two and a half years in, I started dictating voice memos – ideas, fixes for blog posts, marketing thoughts, and then, gradually, everything. The memos fed a Notion board I’d bought a template for, and I’ll be honest: the board went nowhere. But the habit did something the board couldn’t. I was capturing what I knew, and thinking out loud is how I form ideas. By the time a tool arrived that could actually use that stream, I had months of practice at externalizing my own head.

Meanwhile the projects multiplied. I started a blog to make things easier and discovered the content wasn’t good enough yet, so the thing that was supposed to save work became more work. TikToks for the hobby blog, scripts, captures – piecemeal, chaotic, motion without progress.

The December agent, briefly

By late 2025, everyone was talking about agents – AI that does things instead of chatting about them – and I heard “vibe coding” often enough to feel behind all over again. I looked into it and stalled on an honest objection: I didn’t have anything I wanted to code. (If I’d answered that objection with Claude Code then, this post would be a year shorter. I didn’t.)

What I did have was paper. The plan: go digital – scan everything, name everything, tag everything, put it on a shared server my husband and I could both search. In December 2025 I tried to build an agent to do it, and days disappeared into a task a person could do by hand. It halfway worked, once, and I never used it again. The full story lives in How I built an AI that actually knows my work.

I bring it up for what it did to my head. The flop seemed to confirm that the interesting stuff – the automations everyone was posting demos of – sat on the far side of skills I didn’t have. I was wrong, and in fairness the tools really were worse then; today I’d point Claude at that folder and describe what I want in one sentence. But I believed it for months, and believing it is part of what treading water feels like.

Switching to Claude

On March 2, 2026, I signed up for Claude and used it hard from day one – the chat, and Claude Cowork for working with files.

My first delight was almost embarrassingly small: Claude could write into an actual Word file. I could open it, edit it, hand it back, and it would revise – a real document going back and forth, not text trapped in a chat window. I burned through my credits doing exactly that. Then I did what everyone online said to do and made a separate project for each kind of work, and found the trap: my voice guide and my processes lived as copies in every project, each copy drifting on its own, no project aware of what the others had learned. Everyone kept saying Claude has memory now – but exactly where it seemed like it should remember, it didn’t, and nobody could tell me precisely what was remembered where. One big project instead of many worked a little better. Chat with projects, better still – and then a new worry arrived: my working files now lived on Anthropic’s servers, and getting them back out was not obvious.

Plenty worked anyway. I built an online course in those weeks – videos, landing pages, an email sequence – and Claude could hold all of it in view at once and catch me repeating myself from piece to piece, or wandering onto a tangent I’d already covered elsewhere. That higher-altitude audit of my own content was something ChatGPT had never given me. But the limits were the same limit wearing different clothes: sessions starting from near zero, my voice dropping out in handoffs, quality sagging as conversations got long. A smarter assistant, and I’d lost the thing I valued most in the old one: being known. The full comparison, including who should stay in Cowork, is in I tried Claude Cowork, switched to Claude Code.

Claude Code, and the weeks after

On April 21, 2026, I followed a YouTube walkthrough – Nate Herk demonstrating Andrej Karpathy’s idea of giving an AI a personal wiki – and ran my first Claude Code session. In that same session I created the Obsidian vault that now holds my working memory: notes, decisions, projects, all plain files on my computer that Claude reads before we work.

Then things happened at a pace I hadn’t experienced before. Two weeks in, I built a small keyword-research tool in an afternoon – me, the person the file-renaming agent defeated. A publishing pipeline followed, then client work finished faster and more complete than I could have managed alone. My honest assessment, in my own words: “I am leaps and bounds further along than when I was just working in the chat version of Claude or some other whatever AI tool in a project with project files.”

When I described those weeks afterward, the sentence that came out of my mouth was that I’d 10x’d my output. I absolutely hate that phrase – it’s the vocabulary of people I’ve promised not to sound like. So here is the version I can stand behind: I still can’t measure it, but it feels like a lot.

The feeling is the real product

Claude Code gives you speed. But mostly it gives you the tool and the feeling that you can do this – in my words at the time: “yes, I’m going to be able to automate this and do this.”

The feeling has a mechanism, and it works like soup. When you’re already making a pot of soup and you decide to make twice as much, it doesn’t take twice as long – a bit more chopping, a bigger pot, and you end up with twice the soup. That’s what a working system does to your ambitions. Teaching people about AI was never on my list – not because it wasn’t interesting, but because the idea had nowhere to land; there was no spare capacity for it to even feel possible. Now the pot is already on the stove, and teaching is one more thing I add in. New ideas stopped arriving as new burdens. That, more than the speed, is what actually changed.

That feeling is what three years of busy AI use never produced, and it’s the thing I’d actually like to hand you. Not my setup – the tools will have changed by the time you read this anyway. The feeling transfers.

If you’re where I was – capable, busy, quietly convinced everyone else got a head start – the honest news is good. I felt behind on day one at 48. Today I appear to be the most advanced AI user among the people I actually know, and some of them are senior people whose entire AI life is the Copilot that came preinstalled on the work laptop. (I consume a lot of content about AI, which skews anyone’s sense of what “everyone” is doing. The people in the demos are not the people in your office.) You haven’t missed anything permanent. I treaded water for the better part of three years and caught up in a couple of months, and I wasn’t smarter in April than I was in March.

One caution, so this doesn’t turn into the hype I’m allergic to: I can’t promise you my slope. I don’t know your work, and I can’t rerun my own three years to find out whether the treading water was wasted time or the training. None of us gets that experiment.

What I can tell you is where the door was for me: a tool that works in your real files, plus a habit I’d accidentally spent years practicing – writing down, or saying out loud, what you figure out, so the next session starts warmer than the last one. Start tiny. Pick one folder and one afternoon, and let the AI read before it writes. The first hour is mapped in Start Here.