I built an Anki cloze generator that isn't allowed to rewrite your notes
An Anki cloze generator where every answer is an exact substring of your own fact, checked by code. Live test on 20 Oct: 10 of 12 orders done, 2 refunded.
Making cloze cards is the boring part of Anki. Asking an LLM helps, but it likes to "improve" your wording, and then you're memorising its sentence instead of the one you checked.
So Recall Pack has one hard rule: the answer on every card is a substring of your original fact, character for character.
How it works. You paste 1–20 facts (12–240 characters each, 12 KB total) as JSON. The model (claude-haiku-5-5) doesn't write cards. It only returns the quote and the start/end of the span to hide. Code checks that the quote matches your fact exactly, then builds {{c1::...}} itself. If the quote drifted, you get no cards and the order goes to refund. Output: TSV for Anki desktop, JSON, HTML.
Real output from the test: A {{c1::stack}} removes the most recently added item first.
What the 20 Oct test showed. 12 orders through the live pipeline, 10 done, median 5.4 s, max 9.5 s, about $0.00045 of model cost per order. Before that day's fix, every paid order failed: Haiku sends a thinking block before the answer and my parser expected text only. Turning thinking off made it worse (3 of 12 packs came back as prose around the JSON), so I kept it on and filter the block.
Honest weak spots. A 240-character repetitive fact got copied with a change (a direct replay of the same prompt dropped one repetition), so that order went to refund. Sometimes the hidden word is useless (a verb, "различает"). The topic filter is crude: "patient" as plain vocabulary gets rejected. The input is JSON, not a plain list. No medical, legal or financial material.
$4 one-time, no account. Free sample and structure check before you pay. Texts are deleted after 24 hours.