the anti-rot manifesto

i noticed it in my own hands first.

i'd type the first half of a function and stop. not because i was thinking. because i was waiting. there's a half-second where the cursor sits and the completion hasn't arrived yet, and in that half-second i wasn't generating anything. i was receiving. i've started calling it the copilot pause, and once you notice it you can't un-notice it: the moment the thinking leaves your head and moves into the infrastructure.

that's the whole thesis of this series, in one gesture. everything else is footnotes.

the number

this isn't a vibe. it lines up with a randomized controlled trial i keep coming back to: learners split into ai-aided and unaided groups, tested weeks later. the ai-aided group scored meaningfully worse — the gap was widest in technical subjects — and the part that stings: prior fluency with ai tools did not protect anyone. knowing how to prompt well didn't stop the forgetting. i'm going from memory of a paper here, not a link i can hand you, so take the shape of the result, not the decimal points.

the vocabulary i use for it in my own notes is cognitive offloading and borrowed competence — the illusion of mastery without the synaptic wiring. you feel like you learned it, because you were in the room when the answer appeared. you weren't the one who produced it, so nothing got built.

two kinds of rot, one shape

people usually mean feeds when they say brain rot, and they're right. i had 4,842 videos in a watch later queue, which is what passive consumption looks like when you keep the receipts. but the feed is the older, dumber version.

the new version is a tool that thinks for you, and it's harder to see because it looks like productivity. summarize this. explain that. write the function. draft the email. every one of those is fine on its own. together they're a habit of receiving, and the shape is identical to the feed: input arrives, you consume it, nothing in your head changes.

so this isn't a post about ai being bad. i use models all day; most of the code in skia is written by agents. the question is never whether to use the model. it's where the thinking happens. if the answer is "in the model," you're renting a skill. if the answer is "in me, with the model pushing back," you're building one.

active notes

the alternative has a boring name: notes. but not the kind you take. the kind you work.

a note you wrote yourself, linked to three other notes, revised twice because the second draft was wrong — that's a rep. the muscle got used. a summary a model wrote for you that you filed and never opened is a receipt. it proves you were present.

the difference isn't the tool, it's the direction of effort. writing is thinking, and revision is the real thinking. the graph — notes linked to notes, a shape that changes every time you add one — isn't storage. it's the place the thinking happens, and it's the only memory that's actually yours, because you put every node there or you approved it.

that last clause is the product.

what skia is

skia reads what you write, shows you what it saw with the lines it saw it in, and ends on one move you can make today. it's not a notebook — it's the long-term memory for an ai partner, so here is plainly where that memory lives. your notes and everything the app knows live in our database (spacetimedb). it is the one store of record: there is no second vendor holding your data, and no redis or key-value store. your device keeps a copy so the app opens fast and offline; deleting that copy costs nothing, it re-syncs. search, the ai and media need the network. when you use an ai feature, the text it needs goes to google's gemini to produce the answer and the search index. we do not train models on your notes. voice, calendar, web search and the code sandbox reach their providers only when you turn them on. you can export everything and delete everything; deletion erases every table, except stripe's legal payment records and anonymous aggregate counts. analytics are first-party counts of feature names, never note content.

and the partner is built around one sentence i hold every feature to:

it should never hand you the solution. it should help you get there.

concretely: your notes are the memory, and the ai treats what it remembers about you as a hint to check against them, not as truth. when you ask it to write — a note, an edit — it writes, because you asked. when it has an idea of its own — a link between two notes, a distillation — it proposes, and you keep it or you don't. the confirm step isn't ux friction i forgot to remove. it's the rep.

and it asks back. a suggested link comes with its reason and a checkbox — these two look related; are they? — not as a line already written into your note, because the question that makes you check is worth more than the answer that lets you skip checking.

the manifesto, short version

1. receiving is not learning. if you didn't produce it, you don't own it.
2. the tool should make you sharper, not softer. the success metric is people saying "it makes me smarter," not "it makes tasks easier."
3. the ai proposes, the human approves. nothing gets written into your notes without you.
4. friction on purpose. capture in under two seconds; synthesis and deletion deliberately slow.
5. you can leave with all of it. export everything, delete everything, and deletion erases every table.
6. no attention farming. we use your data to serve you, not to keep you scrolling. every reward should leave you sharper, calmer, more in control — the opposite of a feed.

what this series is

three posts. this one is the argument. the second is the only protocol i have real numbers for — 4,842 videos, cleared in about 25 minutes for $0 with all of it kept — and why that's a thinking practice and not a tidying one. the third is the design law that falls out of everything above: a partner, not a repository.

none of this is a return to paper. i tried paper. paper doesn't have backlinks. it's a return to the part of the loop where your head does the work — with a machine that's strong enough to be useful and constrained enough to stay out of the way.

the pause is the tell. next time you feel it, notice who's thinking.