I keep starting posts and abandoning them. Today I'm publishing one before I can talk myself out of it again.
A while back I posted three pieces in a row and then went quiet for months. Not because I had nothing to say — because I had too much, and none of it felt finished. So this one is the opposite: less polish, more honesty. I want to write more often, and the only way I know how to do that is to stop trying to make every piece a definitive take and start treating them as diary entries that other people might find useful.
Here's where I am.
How I got here
I started programming around 14. The usual stuff — Python turtle graphics, W3Schools tabs open at 2am, Geeks for Geeks for anything I couldn't search my way out of.
The first AI in my workflow was Tabnine. A VS Code extension that autocompleted slightly more than the editor's built-in suggestions. At the time it felt like a parlor trick. Looking back, it was the leading edge of everything that came after.
Then ChatGPT. In college, a few friends and I split a Plus subscription — everyone chipped in a fraction of the $20 — and we built a website with it. That site, if you opened it today, would politely ask to be put into witness protection. But it taught me what it felt like to ship as a team, with an AI sitting at the table.
By my fourth year I'd switched to Cursor. Cursor was the first tool that felt less like a chatbot and more like a coworker — I'd write, it'd refactor, I'd read the diff, push back, repeat. It was my daily driver for over a year and I still respect it.
Claude showed up later. I saw a reel mentioning a newer model that was supposedly better at code, jumped on the free plan, and ran it alongside Cursor. For a long time I'd just top up pay-as-you-go, watching my monthly spend creep from $20 to $50, occasionally to $80. Eventually the math became absurd and I bought Pro. Then the 5x. Now the 20x. I'm a slow learner about my own habits.
A lot of what I build today — including the twenty-odd agents we run in production for kAInet — is powered by Claude.
The thing that actually changed how I work
Tools are the easy part of this story. The harder, more useful part is what changed in how I plan and test.
For a long time I worked the way most people do: have an idea, start building, fix things as they break. I called it iteration; in practice it was optimism. When I started using agents seriously, that approach stopped scaling. Agents are confident in places they shouldn't be. Edge cases I would've felt my way through in a hand-written codebase started slipping past me because I'd handed too much to the agent before I'd really thought it through.
Two things shifted that.
The first was a short Matt Pocock video on the AI Engineer channel that pushed me deeper into test-driven development with agents. (It's only a few minutes long — quietly one of the highest-leverage videos I've watched this year.) I'd already been doing bits of TDD, but watching someone articulate it as a discipline reset the whole loop for me. Tests aren't a chore you write after the fun part. They're how you let the agent move fast without you having to babysit every diff. TDD isn't a fit for everything I build — agentic workflows surface edge cases no test suite will catch — but as a discipline it's bought me back more time than almost anything else I've adopted this year.
The second was a skill called grill-me. I tried it on a project where I had a vague brief and no real outcome in mind. Six hours in, it had asked me over 300 questions. I have never been interrogated by software so thoroughly. By the end I had a plan with no soft spots — every assumption named, every gap surfaced, every "we'll figure it out later" pulled forward. The plan isn't a chore you do before the fun part. The plan is the work; the code is the cheap part.
The framing Matt uses — that what you're building with the agent is a shared language, a shared understanding — is the most useful single sentence I've absorbed about working with AI this year. Once that shared understanding exists, anything you build on top of it has a foundation. Without it, you're just gambling with someone else's compute.
What's currently in my workflow
A few things have earned a permanent spot.
Claude does most of the heavy lifting — planning, building, agent orchestration. Cursor still gets used for tighter, hands-on edits where I want to feel the code. Context7 sits alongside both for documentation lookups, so I'm not handing the agent stale memorized APIs. And a quiet but important habit: I still read the code. Not all of it, not line by line, but enough that I know what I shipped. I also write my own edge cases — the agent will not invent the failure mode that bites you in production. That part is still on you.
I lean on a few voices to keep me honest about what I'm missing. The AI Engineer channel on YouTube has been a steady source of "oh, that's how other people are doing it." Theo's rants are entertaining and, more often than I'd like to admit, correct.
What I haven't figured out
I have access to Tinker by Thinking Machines Lab and haven't touched it. Fine-tuning isn't urgent for me right now — my work is mostly agentic, and the agents we run are tuned by prompt and tool design, not weight updates. But "I'll get to it" is doing a lot of work in that sentence.
Front-end is a similar story. I'm a backend person by instinct, and AI tooling has lowered the floor enough that I can now ship UI I'm not embarrassed by. That doesn't make me a front-end engineer. It just means I've stopped using "I'm a backend guy" as a reason not to try.
Why I'm writing this
I want to do this more often, and I don't want each post to be a polished essay. If something I'm trying — a tool, a workflow, a way of thinking — saved me a week, I want to write it down while it's fresh, in case it saves you a day.
If you're earlier in this journey than I am: the tools will change every six months. The discipline of writing a real plan before you write a single line of code will not.
If you're further along: tell me what I'm missing. I'd rather hear it from you than from a production incident.
More soon. Hopefully before next year.