Mole 作者 Tw93 分享 AI 時代如何打造慢產品

It has been a little over two months since I wrote about taking Mole from a CLI to a Mac app. In that time Mole went from 1.13 to 1.18, the Windows version shipped, and the open-source repository quietly passed 73K stars.

Looking back, most of my time did not go into new features. It went into things that look like a bad trade to an engineer: installing 850 Mac apps one by one and then uninstalling them, resending about 400 license emails by hand to buyers who never activated, redrawing dozens of icons over three rounds, and rewriting the interface copy line by line until someone opening Mole for the first time could follow it.

Programmers naturally like work that scales, so whatever the task, we look for a way to turn it into a tool, a script or a model call so we can be lazy. These two months left me feeling a bit different. Code keeps getting cheaper and anyone can quickly build something that runs, so what sets products apart comes down more and more to the details you only learn by doing the work yourself. The good news is that with AI, chores that used to take a small team can now be finished by one person.

Software engineering has no tidy standard

To take the guesswork out of uninstalls, I kept installing Mac apps, launching them, letting them add their login items, then removing them and checking what was left behind. I thought 100 mainstream apps would be plenty, but the list is at 850 now. Some bundle IDs are not even reverse domain names, and a folder with a similar name proves nothing about who owns it, so the only way is to install each one and check it.

My Mac during one of the testing rounds.

The first round of more than 300 apps gave Mole 13 kinds of reusable rules and more than 90 precise cleanup paths. It also fixed more than 30 places I had marked as safe to clean that should have been left alone. Those were as valuable to me as the new leftovers: a cache can be rebuilt, the user's own files cannot, and data shared by several apps should never be wiped just because one of them was uninstalled.

I wrote up the leftover paths and uninstall steps for every app as free guides, so you can follow them by hand without buying anything: mole.fit/tested-apps

You only know an AI tool once you have used it

People asked for AI tool cleanup long before Mole had it. I used to treat coding session history as a valuable asset and only cleaned real caches. Then a single user's opencode.db grew to 62 GB and they asked how to clean it, so I tested more than 30 AI tools one by one, and this time I signed in to every one I could and used it once in a real project, because a lot of the data only appears after real use.

Some of it was surprising. With all 13 desktop apps open at once they took about 23 GB of memory, and one of them started 41 processes on launch. Some put a root-owned symlink in /usr/local/bin that turns into a dead link once the app is deleted. One command line tool, after I ran it once, downloaded its new version without asking and edited my .zshrc.

AI Cleanup & Care in Mole.

That changed one of my earlier assumptions. Sessions really are valuable, but old sessions and worktrees you are done with really do take up space. Mole's AI cleanup now lists sessions by the retention period you choose, leaves them unchecked by default, and keeps a recovery copy before deleting. None of that came from an architecture diagram, it came from using the tools until they misbehaved.

Resending 400 emails

Over China's National Day holiday I went through about 400 Mole orders from the May launch to the end of September that were bought but never activated, and resent the keys one by one, grouped by purchase month, country and language. Not a promotion, more like a message to a friend: a reminder not to forget it, and no rush if they are not using it yet.

Replies kept coming in. Some had bought it just to support the CLI and still preferred the Terminal. One had bought a second copy as a Christmas present for a family member. Quite a few had simply typed their email wrong, gmai.com or gmail.con, and never got the first message.

In the dashboard all of these people had the same status: not activated. It took reading the replies one by one to see the difference. So after payment the site now shows the key right away, and clicking Activate there opens the installed Mole through a deep link and fills in the key. It never activates on its own, so the last click stays with the user.

Polish that looks like wasted effort

Paul Graham wrote in "Do Things That Don't Scale" that Wufoo sent handwritten thank-you notes to early users. Now that code is hardly a moat, that kind of patient, unscalable work is worth even more. AI makes building a passable interface easy, but it will not decide on its own to install 850 apps, redraw an icon three times because users could not tell what it was, or rewrite a thank-you letter until it stops sounding like a marketing email.

Some of Mole's icons, before and after the second round.

If you are building something in this fast AI era, leave room for the slow parts. The product is often hiding there.

The longer version of this post is on my blog: https://tw93.fun/en/2026-10-26/mole-product.html, and Mole for Mac is at https://mole.fit