Why
Why I built it
I wanted an AI tutor that adapts to how I answer without corrupting Anki review history if a chat drops, the model forgets context, or a request retries.
What I built
What I built
- The AI teaches and grades while Anki stays the source of truth for cards and spaced repetition; the card-by-card flow grades and fetches the next card in one tool call.
- A lesson journal logs sessions, cards, and first-answer evidence so progress carries across conversations; follow-up corrections never overwrite the initial grade.
- Lesson completion previews writebacks before commit and marks pending submissions uncertain so interrupted requests are checked rather than replayed.
- Works with standard MCP clients, including background-tunnel setups, with teaching preferences passed at lesson start.
Engineering
Problems I worked through
- Separates exposure from recall: only cards with recorded answer evidence count toward final grading.
- Because a card might change on another device during a long lesson, due state and daily quotas are re-checked before writeback.
Role & AI
How I split the work with AI
My daily study routine defines the requirements; I work with coding agents to refine the tools and lesson journal so teaching feels natural without risking review state.
Outcome
Outcome
Used daily in my own language study, with both card-by-card and persistent lesson modes running while Anki stays authoritative for review scheduling.