I think faster than I can talk. For years that meant my ideas lived on Post-it notes, and I could only keep one or two projects moving at a time.
AI changed that. The way I work isn't that different from anyone else's, but the tools, skills and workflows I've set up work like a superpower. I capture ideas as fast as I have them. I keep many projects going at once, each in its own chat. Reminders keep me on track. And I learn much faster than I used to.
I'm also a constant tinkerer. If something can be automated, I want to automate it, whether it's data, tech, IT or something I'm building.
A year ago I didn't work like this. Six months ago I worked differently than I do now, and so did three months ago.
This is my setup as of November 2026. If you're reading this in a few months, it's probably already changed.
The short version
- I treat AI like extra roles on a team: an engineer, a report writer, a researcher, a tutor. My job is to decide what gets built, check it and own the result.
- Anything I do more than twice becomes a skill: written instructions and examples that Claude follows the same way every time.
- AI writes most of my code now. I read along, ask why, and make the architecture calls myself.
- Nothing goes to anyone until I've checked it against something I already trust.
The stack
- Claude for work, and Claude Code for building and shipping code.
- Custom Claude skills for recurring analysis, reporting and development work.
- Muse as my personal layer. It's connected to my email, works almost like a to-do list, and holds small apps I built for spending, workouts and training. Claude is where I do the serious building and professional work, like this website.
- Python and SQL (SQL Server, PostgreSQL) under almost everything.
- GitHub for every project, with tests and scheduled runs on GitHub Actions where it fits.
- Supabase, Render, Cloudflare and Vercel for databases and hosting.
- TickTick for tasks and Monday.com for project tracking.
A typical day
My mornings start on paper. A Python pipeline I built pulls my notes from the day before, my calendar, reminders, market closes, scores and the weather into a newspaper layout, adds a crossword it generates, and prints it at home. I read it before I pick up my phone.
Those notes come from Muse. I keep a side chat there where I record random thoughts as they come up during the day. Each one gets logged to a GitHub repo that works as my living journal, and the pipeline pulls yesterday's entries into the print. So every morning I review what I was thinking about yesterday. It gives me topics to journal on and gets my head into the day.
During the day, an AI assistant checks job boards at set times and sends me only new postings that match the roles I want. When one fits, I hand it the posting and it tailors my resume to that job before I apply. I still read every line before it goes out.
I keep several chats and projects going at once, grouped by what they're for. Muse reminds me about the projects I'm working on in Claude, and Claude sends me reminders through the Claude Duet app, so nothing slips.
Workflow 1: Reports that run themselves
The problem. Monthly revenue, bi-weekly intake, quarterly shipment counts. Same queries, same charts, same write-up every time, and a lot of copy and paste in between.
The setup. I turned each one into a Claude skill. A skill holds the SQL, the chart spec, the output format and the rules for what done looks like. My monthly revenue analysis comes out of one run as Excel, Word, PowerPoint and HTML. A changelog skill turns merged GitHub pull requests into a Word document for people who will never open GitHub. For a weekly meeting agenda skill, I made the last completed agenda the memory: Claude reads what we covered last time and builds from it, so there's no separate notes file to keep in sync.
When my check finds a problem, I fix the skill, not the output.
The result. The manual work turned into a check and a send. That didn't shrink my workday. It moved the goalpost: my time now goes to learning, thinking, and working out better ways to process and review data.
Workflow 2: Internal tools in an afternoon
The problem. At my last job, a dental lab, when the floor needed to follow a new training plan or quality check, the usual answer was a spreadsheet and a reminder email. Neither one sticks.
The setup. I build web apps with Claude Code in Python, FastAPI and PostgreSQL. I make the big decisions and Claude writes most of the code. One example is a barcode scan station that technicians use on iPads with Bluetooth scanners. I chose to log every scan in a new, clean database and write it through to the old lab system at the same moment, instead of building a two-way sync. The old system stayed the source of truth while the new one grew around it. Before trusting it, I compared the first real scan with what the old system writes on its own, and they matched. When the confirm screen felt slow, I had Claude find the slow step and move it to the background until it felt instant.
The result. When the floor needs something new, I can spin up a production-grade app that reinforces it in an afternoon or overnight. The biggest build takes a submitted appliance prescription through check-in, design, QC and nesting onto the 3D printers with no human touch until after printing.
Workflow 3: Side projects as a lab
Side projects are where I try new tools before I need them at work.
- Job alert tool (open source, on my GitHub). Job aggregators re-list postings hours or days late, so this reads company job boards directly every hour (Greenhouse, Lever, Ashby and Workday), keeps only postings it hasn't seen, filters by title, location and age, can score each one 0 to 100 against my resume with Claude, and sends the best matches first by email or phone push. The test suite runs with no network. It's running now and sends me new matches as they post.
- Triathlon training app. It merges my Garmin and Apple Watch data into one live dashboard. A planner script builds my next session from recent workouts and my training plan, then pushes it to Zwift, the treadmill or a phone reminder.
- Printed daily brief. The newspaper from my mornings, above.
- Small apps in Muse. A spending tracker that refreshes a few times a day, a workout log for lifting sessions, and a fitness dashboard with a rolling three-week training plan that updates weekly.
I let Claude write most of the code on these, but I follow along and ask questions until I could explain every part. That's the point of them.
Workflow 4: Learning with AI
The problem. I wanted my own deep dives with good charts, so I built an AI research team to do the digging for me. My topics go everywhere: how airplanes fly, engines versus EVs, past industrial revolutions versus AI, and a lot of markets and macro.
The setup. A deep-dive skill writes a paper on the topic first. Then it quizzes me and has me teach the topic back, and it points out what I got wrong. Each topic gets its own chat in one project. Depending on the topic, I also have it make me a study guide, a short video or a podcast.
The result. The papers that hold up become posts here, like my deep dive on oil and the 10-year Treasury.
Try this: talk to your AI in the car. If I'm learning how airplanes fly that day, I spend the drive in a conversation about it, asking questions out loud. It's the habit I recommend to everyone.
When AI gets it wrong
When I first built my revenue forecasting skill, the department breakouts looked too even. One product line had only launched in August, and the model forecast it as if it had been selling all year. It was easy for me to spot, and the model missed it completely.
Mistakes like that happen less now, but it changed what I review. I start everything from first principles. I need to understand the key levers and drivers well enough to know when something is off. Then I write tests that trigger those errors on purpose, instead of relying on a spidey sense, which isn't real and doesn't scale.
That's the trick with AI. It's a muscle you work every day, and you get better at noticing when a model or a skill starts drifting. It's like being a boat captain, constantly making small corrections to keep the ship on course.
If you're starting out
- Pick one task you do every week and write it up as a skill or saved instructions this week.
- Give it a real example of good output.
- Check the first few runs line by line, and fix the instructions, not just the output.
- Build one small thing for yourself. Personal projects are the cheapest place to learn.
- Ask the AI to explain its code until you could explain it to someone else.
If any of this sparked an idea, or you want to compare setups, I'm always up for that conversation. I'm also looking for my next role, so feel free to reach out.