The AI tools I actually use every day
# The AI tools I actually use every day
Every few weeks another list lands with twenty AI tools you supposedly cannot work without. I have signed up for a depressing number of them. Most lasted under a week. The excitement carried me through the demo and the first afternoon, and then the tool sat untouched while I went back to how I already worked.
So this is the opposite of a top twenty. It is the short list of AI tools that survived a year of daily use on my machines, and the test each one had to pass to stay.
The test is disappearance
A tool earns its place when I would feel its absence the same day. That is the whole test. If it vanished this morning, would my work get slower or worse before dinner. If the honest answer is that I would notice in a month, or never, the tool was decoration.
By that test my keeper list holds five entries. My install history holds dozens. The five that stayed share one shape. Each does a single job, does it the same way every time, and slots into work I was already doing. The ones that died usually asked me to reorganise my day around them first.
A writing model I read skeptically
The tool I lean on hardest is a large language model for drafting and tightening text. This is also the most oversold category in the whole space, so let me describe the narrow way I use it.
It gets me from a blank page to a rough draft in minutes. The blank page used to cost me real hours, and that gap is the entire value. I read every line it produces before anything ships, because it pads, and occasionally it states something false with total confidence. Catching that is my job. I never hand it the thinking and I never ship its output unread.
It is a paid, metered service, and at the volume I write the bill stays small. If it disappeared tomorrow I would feel it within the hour, which is the strongest endorsement I can give any tool.
One voice I stopped auditioning
The second keeper is a text to speech voice. I run faceless video, so the narration voice carries the whole production. Nobody ever sees a face. The voice is what gets judged.
I auditioned plenty of options. Most sounded convincing for a sentence and drifted over a ten minute script, a wobble here, a strange emphasis there, until the read felt off in a way a viewer could hear without naming it. The one I kept holds steady across a full narration, take after take, with no babysitting. Consistency decided it, ahead of raw quality. It is metered and paid, and the per video cost is small enough that it never enters the decision of whether to make something.
Image generation on my own graphics card
The third keeper runs entirely on the card in my own machine: local image generation for thumbnails, simple visuals, and rough concept art.
Honesty requires saying the hosted services often produce sharper images. Mine survived on economics. I generate images in volume, and locally the marginal cost of one more is a little electricity. Run anything hundreds of times and that arithmetic decides everything. I also skip other people's queues and never wonder where a prompt went.
If I needed one image a month I would use a hosted service and forget the whole subject. Volume is the only reason this earns a slot.
A narrow agent I still check
The fourth keeper took the longest to trust: an agent that can read my files, run small commands, and chain a few steps together without me clicking through each one. For a long time I filed these under demo toys, impressive on stage and useless on a normal Tuesday.
What changed my mind was wiring one into the tools I already use instead of working inside somebody's sandbox. Now it handles the multi step chores that are too small to be interesting and too frequent to ignore. Renaming a batch of files. Pulling one number out of a log. Kicking off a job and confirming the result.
Two habits keep it useful. I hold the scope narrow, small tasks where a glance at the result tells me whether it got things right. And I read what it did, every time, because it does go off track, and I want to catch that before it touches anything that matters. Trust here is earned slowly and lost in one bad afternoon.
Plain scripts, barely AI at all
The last keeper hardly qualifies for the category, which is exactly why it survives: ordinary scripts running on a schedule. They move files, rename things, stitch steps together, and fire overnight without me.
Here is the uncomfortable pattern I keep meeting. A large share of what people reach for a fancy AI product to do turns out to be a five line script. The model sometimes helps me write that script faster, and I will take the help. The thing running in production, though, is dull, predictable code that behaves identically every night and never surprises me. The most reliable component of my whole setup has no intelligence in it whatsoever.
The graveyard sorts into four piles
I keep a much longer mental list of tools I dropped, and the reasons repeat so consistently that they are worth more than any recommendation.
Pile one is novelty. Tools where the excitement was the entire experience. A text to video toy, a summarizer for every page I opened, a note system that promised to think alongside me. All genuinely clever, and none solving a problem I had on an ordinary working day. Fun wears off on a schedule.
Pile two is friction. Tools that produced good output only after so much coaxing and tweaking that the effort ate the savings. One image tool gave me incredible control and I came to dread opening it. When the cost of getting a result never drops, the math quietly fails.
Pile three is islands. All in one platforms that wanted to host my writing, my images, and my planning in one dashboard. Each one asked me to leave the pipeline I already had, and none could talk to it, so every session meant copying material in and out by hand. A decent tool that plugs into everything beats a great tool that lives alone.
Pile four is defeat by something dumber. An AI scheduler lost to a recurring calendar reminder. A summarizer lost to the habit of reading the thing. A clever file organizer lost to a sane folder structure and one small script. The simpler option was predictable, fully understood, and never had an off day.
Time, fit, trust
Three questions decide keep or cut, and every candidate faces them once the honeymoon ends.
Does it save real time every single week. If I cannot point at the hours, it goes.
Does it fit the pipeline I already run. A tool that requires leaving my setup and shuffling material around by hand spends its savings on friction before I see any.
Can I trust it. Meaning I can check its work quickly, it behaves the same way twice, and it will not quietly break something while I am elsewhere. A tool I have to worry about is a cost wearing a savings costume.
Fail any one of the three and the tool goes, however good the launch video was.
Chasing tools is comfortable
The warning I most need to hear myself: evaluating new tools is one of the most comfortable forms of procrastination available. It feels productive. You are reading about AI, testing things, tuning your setup. Much of the time you are avoiding the actual job with extra steps. I have burned entire afternoons configuring something that saves ten minutes a week, and the configuring felt like progress the whole way through.
The people getting the most out of AI right now tend to carry a small kit. They found a few tools that hold up, went quiet, and did the work. That is the entire method I have to offer. Stop collecting, and keep only what you would miss by dinnertime.
For more plain accounting of how I actually run AI day to day, there is more at [xavierfok.com](/).
Get new guides and videos first — join the Telegram channel.