The one prompt change that fixed extraction quality
Posting a quick build-note because the lesson surprised me.
First version of LinguaMind's vocab extractor was a single LLM prompt: "given this text, return the useful vocabulary." Output was fine. Generic-fine.
Adding two things to the prompt 10x'd it:
- The learner's CEFR level (A1 / B1 / C1 etc.)
- Three concrete words the learner already knows comfortably
The three known-word examples anchor the model far more reliably than the abstract level label alone. The same email gives wildly different correct answers for an A2 learner vs a B2 learner, and only the level + known-words variant reads as "uncanny how well it picked the right ones."
This is now the same shape every MeDo LLM-plugin call in LinguaMind uses: pull the user's state from the persistent storage plugin, condition the prompt on it, write any new state back.
If you're building on MeDo, this pattern is reusable everywhere.
Log in or sign up for Devpost to join the conversation.