GitTok 1.1: your own writing guidance, and a feed that learns better¶
GitTok 1.1 lets you tell the card writers how you want your posts, makes pairing a phone one simple step, and fixes a set of ways the feed could learn the wrong thing from what you do.
Then say "update this GitTok directory" in a Claude Code session there.
Your own writing guidance¶
The mix and the rules decide how many posts of each kind get made. Now you can also say how to write them. local/guidance.md in your GitTok directory is free Markdown for the writers:
# How I want my cards
- Keep hooks under 12 words.
- I am new to Clojure: explain its syntax when a card shows Clojure code.
- No pop-culture references. Puns are fine.
Every writer reads it from the next run on. Or just ask Claude ("keep hooks short") and it edits the file. See Steer the feed.
Pair with a code, everywhere¶
There is now one way to pair a phone, and it works the same in the browser and in the app on your home screen:
- Open the feed address on your phone. Claude gives it at the start of each session, or ask: "what is the feed address?"
- Tap Pair with a code.
- Tell Claude the code: "pair code X6J2-QG".
The QR code page and the pairing link are gone. No page or link holds the pairing token any more, so there is nothing to leak in a screenshot or a chat. See Get started: pair your phone.
A feed that learns better¶
A review of the learner found several ways it could learn the wrong lesson. All are fixed:
- One short look no longer turns a read post into a skip. Read a post for 20 seconds, come back to it for half a second, and it used to count as skipped. Now a post is a skip only when every look at it was short, and an answer or an opened panel always counts.
- Watch time is counted once. A comic or a chat you came back to could count its play time twice.
- The feed shows what exists, in order. The learner now decides only what gets made: the mix of each new batch. The feed shows every post you have not seen, oldest first, and a 👍 or 👎 no longer reshuffles the posts ahead. A post you saw does not come back a week later (quiz reviews still do).
- A 👎 you take back is unlearned, and the story can be told again.
- "New to you" means new when you saw it. The learner could record code as familiar because of the very post that showed it to you.
- A meme template you disliked no longer goes before one you never saw.
- Rebuilding the model keeps what it should.
gittok relearnnow uses the posts' current features, keeps a reset, and keeps what a removed repo taught. - Smaller fixes. No impression loses its score after a clean-up. Quiz answers use the phone's clock, like the other events. Rejected leads are forgotten after the cooldown. A zero in the decay or prior settings is a config error, not a crash.
- Faster. A learning pass reads only what is new.
The model is now trained on cleaner data. The details are in Learning and ranking.