★ Pinned
The week I stopped arguing with my todo list.
After eight years of trying to make GTD work, I gave up. Then I built the thing I actually wanted.
I have an embarrassing confession. I have used — and abandoned — every productivity tool that's been written in the last decade. Things. Todoist. Notion. OmniFocus twice. A bullet journal for an entire year. Three different versions of Sunsama. I've paid for lifetime tiers I never used past month three.
Every time the failure mode was the same: I'd write everything down. The list would become unmanageable. I'd start ignoring it. I'd feel guilty. I'd start a new list. Repeat.
The tool was never wrong. I was using it correctly. The premise was wrong.
Lists are passive. They sit there until you organize them. The tool waits for you to be the system. But I didn't want to be the system. I wanted a system that could read what I'd written, hold the picture, and tell me what to do.
Photo · Maya's desk · 6:14am · the morning Krytz crossed the threshold
That's the bet behind Krytz. The system has an opinion. It captures the messy fragments. It reconstructs the live state. It commits to one primary action with the reasoning visible. The rest stays quiet.
I'm writing this on day 84 of using my own product. I have not opened Things since day 6. The thing is real.
— Maya
Why we will never add a streak counter.
A short note on dopamine, dignity, and what we owe the user.
A friend asked me last week why Krytz doesn't have streaks. Or badges. Or weekly stat emails. "Wouldn't that drive retention?" they said. Yes. It would. I told them we would never add them.
Productivity tools learned a bad lesson from social apps: dopamine works. Reward people for opening the app. Reward them for completing things. Reward them for being "consistent." Watch them check in daily. Watch them feel terrible when they break the streak. Watch them blame themselves.
The job of a productivity tool is not to make you want to use it. It's to help you finish your work and then leave you alone.
If Krytz is making you better at your week, you don't need a counter to remind you. If it isn't, a counter is the wrong fix. Either way, the answer is no.
— Lex
The first time the prioritization was wrong, and what we learned.
A small but real story about override, memory, and the day a user yelled at us.
Three weeks into the beta, a user named Asha sent us an email that began: "Your priority layer is dumb." She had a Q3 deck that the system kept prioritization at #6, even though she'd written about it five days in a row and the board meeting was Tuesday.
We dug into the trace. The deadline signal hadn't shifted because she'd never typed a date. The mention signal was high but the dependency signal was low (no other items pointed at it). And we'd intentionally clamped how much mention alone could lift a score, to prevent jitter.
We were technically right. We were practically wrong.
Two things shipped in the next sprint: a tighter mention curve for items mentioned 3+ days in a row, and the "override · remember" pattern. You correct it once. The model learns.
Asha's the reason we have that feature. We send her a card every quarter.
— Maya
I was wrong about vector databases.
Eight months into the build, I rewrote the memory layer from scratch. A short post-mortem.
When we started, I was convinced semantic memory belonged in a vector database. I'd worked with one at my last job and it was great. Embeddings, cosine similarity, sub-100ms recall. What could go wrong?
What went wrong: we needed time. Krytz's memory isn't just "find me a similar capture." It's "find me what the workspace map looked like at 3:18pm on Thursday, and tell me what changed by Friday morning." That's not vector search. That's chronological history with a time axis. We were using the wrong tool because I had a hammer.
We ripped it out in February. The current memory layer is a sharded, chronological log with a separate semantic index on top of it. Recall is faster. Recall is more accurate. And — the part I didn't expect — recall is more auditable. Every answer points to a specific event with a timestamp.
I should have built it that way the first time. I'll get it right the second time.
— Lex
What I want the first ten thousand users to feel.
Not a metric. Not a NPS. A feeling.
Most product teams write OKRs. We write feelings. The feeling for the first 10,000 users is this: quiet competence. They open the app. The system already knows. There's one thing to do. There's a reason. They do it. They close the app.
If at any point in that loop the user feels stressed, that's a Krytz bug, not a user bug.
— Maya