Last week, without writing a single line of code myself, I built a phone line that picks up when I’m not free, watched a feature go live on a website I run, and set up a system that quietly documents every song I make as a hobby. All of it happened through conversation — describing what I wanted, testing it, and correcting course when it didn’t work the first time.
I’ve been experimenting with AI assistants for a while now, mostly for small automations around my home projects. But this past week was the first time I noticed just how much ground you can cover when you treat an AI less like a search engine and more like a collaborator you talk a project through with — and how little uninterrupted time that actually takes.
Work was genuinely chaotic all week, so most of this didn’t happen in some dedicated block of free time. The website work and the writing got done in twenty-minute stretches on the train to the office. The bigger, more technical project — the phone assistant — was a weekend build.
A Phone That Actually Answers
That weekend project had been sitting half-finished for weeks: a small assistant that answers my phone when I can’t, built and hosted on my AI platform, bacher-ai.com. Building the actual calling infrastructure — the part connecting a real phone number to a piece of software — turned out to be the hardest technical piece of the whole week, and it had defeated a few earlier attempts with older AI models. This time, using a newer, more capable model that had just become available, it came together in a single session — I wrote up that side of the experiment separately in Testing Claude’s New Fable 5 on a Real Infrastructure Problem, and the more product-facing result in Our Voice Agents Can Now Answer Calls.
Once the plumbing worked, the assistant itself came together fast: a voice that greets callers, takes a message, and emails me a summary. The first version had a strange habit — asked who she was, she’d occasionally blurt out the name of the underlying AI model powering her, breaking the illusion entirely. Swapping to a different, more natural voice model fixed that, and made the whole conversation flow far more smoothly.
A Website Written On The Train
My privacy website, prettysimpleprivacy.com, got real attention this week too — all of it fitted into those same train commutes. Two of the site’s Simple Privacy Fix guides went live: one on messaging, and one on VPNs — each a short, practical guide walking through one genuinely simple step a reader can take. Alongside those, three new blog posts went out: Naomi Brockwell’s New Video On Digital Independence, What VPNs Actually Protect You From, and The Most Powerful Surveillance Device You Own. None of it took a sit-down afternoon — it got built in the gaps between checking email and getting off at my stop.
A planned feature to showcase content from privacy creators I trust also came together on the site, though simpler than first intended: reusing a feed I already had, plus short commentary posts like the Naomi Brockwell one, rather than building a dedicated system from scratch.
Automating The Fun Part
Alongside all this, I make AI-generated music as a hobby, under the name DJ MaiTRX. It’s genuinely fun — but every finished track used to mean manually noting the prompt, the settings, and other details for my own records. This week, that entire process became automatic: every time I generate a track now, a small file documenting it gets created and saved without me touching anything. One less chore between me and just making music.
Where I Still Had To Steer
None of this happened on autopilot. The phone assistant needed several rounds of real testing on real calls before it behaved the way I wanted — at one point I made the call to drop a feature entirely (a system meant to detect spam callers) because it simply wasn’t working reliably enough to keep. Other times, fixing a problem — like the assistant switching languages mid-sentence — meant resisting the urge to over-explain what I wanted, and instead simplifying the instructions I gave it.
That, more than anything, is what stood out this week: the AI could build fast and suggest good approaches, but deciding what was actually good enough to keep, and when to change direction, was still very much my job.
A year ago, a week like this — on top of an already demanding work schedule — simply wouldn’t have happened. Instead, I ended up with a phone line, five new pieces of writing across my privacy site, a small feature update, and a hobby automation, all genuinely working, built almost entirely in scraps of time. I don’t think that makes the work disappear. It just changes what the work actually looks like, and how little uninterrupted time it now needs.
Looking back, this was just a genuinely great, productive week — the kind that reminds me why I enjoy tinkering with this stuff in the first place. I’m already full of ideas for what to build next, and looking forward to it.
Key Takeaways
- Building real, working tools with AI is realistic in days now, not months
- It doesn’t require dedicated free time — most of this happened in train commutes and one weekend
- The hardest parts are still the “plumbing” — connecting different systems together
- AI can suggest and build fast, but deciding what’s good enough to keep is still a human call
- Testing on real, live use is where problems actually surface — not in the planning
Photo: SHVETS production via Pexels
This post was written with AI assistance, based on a summary of the week’s Jira issue tracking used across all bacher-ai projects. Reviewed and edited by me before publishing.