Can Claude Code Opus make high quality language learning decks? I tried

Anki Learning Effectively
1

I have been studying Japanese with the Jlab deck for Japanese which aims at teaching Japanese (mostly grammar). It is structured in such a way that each card introduces at most 1 new item: grammar point or word, this strategy is commonly referred to as i+1 or n+1 sentences deck.

I was impressed at how good & smooth learning with this structure is and since I recently received a free Max subscription from Anthropic (I had canceled a previous subscription before their usage limit fiasco), I decided to attempt to create a Jlab-like deck for learning grammar in my target language, Thai :thailand:.

I figured out Claude Code would be the best way to iteratively create a deck. I used a custom CLI tool called tsv-finder to have the AI (namely, subagents) search through a database of 124,000 Thai sentences extracted from native Thai shows by a subs2srs-like tool.

TLDR

Is it possible for a state-of-the-art LLM to create language learning decks following the i+1 sentences design?

⟶ With simple scaffolding (1 CLI just to query for sentences, no programmatic pre-scoring of sentences) and a 2-layer agent hierarchy (lead Claude Code + sub-agents), the answer is a categoric NO.

If more sophisticated scaffolding and a nested agent hierarchy (i.e. sub-agents that can create sub-sub-agents to lessen their workload) were used, the answer is perhaps… but this is not supported in Claude Code for now.

Are large language models capable of writing a pretty good grammar deck with an approximate order of difficulty respected?

⟶ The answer is yes! They absolutely can do a good enough job for the resulting deck to be useful. In my case the grammar progression and the exhaustive showcasing of grammar points through real world examples looks great so I will be studying this Thai grammar deck.

Some stats:

  • Average of 2,587 tokens used per Anki note created
  • Total token usage (in+out): ~3.9M
  • Estimated cost if the Anthropic API had been used: ~$81

Repo: GitHub - tassa-yoniso-manasi-karoto/exp-LLM-mining-grammar-anki-deck-thai

Anki deck: https://ankiweb.net/shared/info/2013132445

2

Which assessment brought you to this conclusion?

they can produce pedagogically useful grammar decks

2 Likes
3

well, a deck that showcases grammar points and their nuances systematically using real world example is useful.

even if not as refined as a human-made deck, the sheer difference of time alone it would take to make one manually with a comparable or superior level of refinement makes it worth it for less popular languages,… or even for popular languages when trying to learn more niche grammar / idiomatic patterns.

The alternative for many languages isn’t a perfect human-curated deck, it’s no deck at all.