The Afterimage blog
How I rebuilt Rewind in 2 weeks with Claude Code
Rewind went dark in December 2025. So I rebuilt it as a private, local-only Mac app in two weeks with Claude Code, on a laptop with 5 GB of free disk space. Here's the story.
On December 5, 2025, Meta announced it was buying Limitless, the company behind Rewind. Two weeks later an update switched off recording in the Rewind Mac app. If you relied on it, your Mac went back to forgetting everything overnight.
I wanted that superpower on my own Mac, with one change: nothing would ever leave the Mac. No account, no cloud, no AI reading my history. So I built it, with Claude Code as my pair programmer, in about two weeks. This is that story.
The laptop with 5 GB left
Here's the part nobody tells you about building a screen recorder: you need disk space to test one. My MacBook had about 5 GB free. Not 5 GB for the project. 5 GB, total.
So the whole build ran in what I started calling “lean mode.” Every milestone ended by deleting everything it had built. Test recordings lived for five minutes and then went in the bin. A build script refused to run if free space dropped too low, because a full disk on a Mac doesn't fail politely.
The constraint turned out to be a gift. When your own disk is nearly full, you feel every megabyte an app writes. Afterimage's whole attitude to storage, keeping a year of history in a corner of your drive, comes straight from building it on a machine with no room to waste.
It also produced the strangest bug of the project. At one point, reading text off the screen suddenly took nearly a minute per frame. Nothing in the code had changed. The disk was simply too full for macOS to keep the files it needed to do the job quickly. We switched to an approach that needs no scratch space at all, and it turned out faster and lighter on the battery anyway.
A plan before a prompt
The most useful thing I did happened before any code: a one-page plan. Afterimage would do exactly two things. It would keep a Timeline of your screen that you can scroll through and search, and it would capture your meetings automatically, with transcripts. Everything else got cut.
Each feature came with numbers it had to hit, measured rather than hoped for. It had to stay light on the processor, small in memory and frugal on disk, and the Timeline had to open instantly and scroll smoothly. Every milestone ended with a test run and a table of results. If a number got worse, we stopped and found out why.
What working with Claude Code was actually like
It's less “the AI wrote my app” and more “I got a tireless senior engineer who never gets bored.” A screen recorder is mostly unglamorous work that has to be exactly right: recovering cleanly from a crash, handling two displays, asking for macOS permissions without confusing anyone, testing everything. Claude Code did that work quickly and consistently, and it was just as happy writing the measurement scripts I'd normally put off.
My job was judgment. I decided what to build and what to cut. I read the numbers and noticed when one looked wrong. I said “no, a real person won't understand this screen.” And I kept the project honest: every outside source we looked at went into a ledger, and nothing was copied from anyone.
Nine rounds of “would a normal person get this?”
The app can take screenshots of itself at different screen sizes, which made design reviews fast. Afterimage went through nine of them. Each was a list of things a real user would trip over: a button that didn't look clickable, an animation that stuttered, a sentence that didn't fit in German. The animated Timeline now uses a fraction of a percent of CPU while it's open.
That loop of build, screenshot, critique, fix is where most of the two weeks went. Writing code turned out to be the fast part. Deciding what good looks like was the slow part.
The numbers I'm proud of, and the one I'm not
From the final test runs on an M1 Pro:
- About 1% CPU while recording a quiet screen
- About 150 MB of memory
- Zero network connections from the recorder, checked every 10 seconds
And the honest one: under really heavy screen activity it averaged about 5% CPU, above the 3% target I set myself. Reading text off a busy screen is real work. I'd rather tell you that than hide it.
What “two weeks” doesn't capture
The core came together fast. Then came the part that turns a project into a product: translating it into ten more languages, getting it signed and notarized by Apple so it installs without warnings, and building licensing that works without a server of our own.
If you're building something similar
- Write the budgets first. “Fast” isn't a requirement. “Opens in under a third of a second” is.
- End every milestone with a measurement. Every serious bug in this project was found by a number that looked wrong.
- Constraints are features. A nearly full disk made a better product.
- Review the product, not the code. Screenshots at real screen sizes caught more problems than any code review.
Afterimage is the result: the best idea Rewind ever had, rebuilt to live entirely on your Mac, for $29 a year.