"AI activities" have a reputation for sounding flat: vocabulary that's slightly off-level, example sentences no one would actually say, the same ten "safe" words every time. That's almost always a prompting problem, not a model problem. Used well, AI removes the tedious part of activity-building while leaving the judgement where it belongs: with you.
Generic output comes from generic prompts
"Make a Spanish vocabulary quiz" will give you a generic Spanish vocabulary quiz. The fix is specificity. A useful brief names:
| Specify | So that |
|---|---|
| Level (a CEFR band: A2, B1…) | the difficulty matches your learner |
| Topic (the real theme: "ordering food") | you get lesson content, not "vocabulary" |
| First language | distractors target the mistakes your learners make |
| Format and length | you get a usable draft, not an essay to reshape |
The difference between "make a quiz" and "make 10 A2 multiple-choice questions on restaurant vocabulary for Spanish learners whose first language is English" is the difference between filler and a usable draft.
Especially on what learners are most sensitive to: register, idiom, and the occasional confidently-wrong answer. Build in a 60-second edit pass. Scan for factual or grammatical errors, check that a native speaker would actually say it, and trim anything off-topic or off-level. The model drafts the boring 80%; you own the 20% that makes it yours.
There's a bonus hiding in that specificity, and it isn't only about avoiding the generic.
Craik and Lockhart showed that how deeply material is processed predicts how well it's remembered, the levels-of-processing effect. Material tied to something specific and personally relevant is processed more deeply than generic filler. So a quiz built around your learner's actual trip to Berlin isn't just less bland than a stock one. It's more memorable, because it's anchored to something real. The specificity you add to beat "generic" is the same specificity that helps it stick.
Craik & Lockhart, levels of processing.
Specific in, specific out. Here's the kind of playable draft that workflow produces.
A specific draft, playable
Match each phrasal verb with get to its meaning.
0 of 8 matched. Tap a term, then its match.
Use AI for the build, not the pedagogy
AI is best at the mechanical work: generating candidate items, conjugation sets, distractors, example sentences. It's worst at deciding what your specific student needs next. Keep that decision human. Pick the format that fits your goal first (how the formats map to learning goals), then let AI fill it.
This is the workflow Lingoken is designed around: describe the lesson, get a playable draft in a real game format, then edit before you assign it, rather than copy-pasting a wall of text out of a chat window and rebuilding it by hand.
Used this way, AI doesn't make your teaching generic. It gives you back the hour you'd have spent typing, so you can spend it making the lesson engaging.
Sources & further reading
- Levels-of-processing effect. Why specific, deeply-processed material is remembered better.
- The CEFR. The level bands to name in your prompts.
Written by
Builder at Lingoken
By day I build, and I love teaching kids to code. On weekends I dance, contemporary and Kathak. Languages are the thing I keep coming back to: I grew up between a few (Sindhi, Hindi, Marathi, English) and I'm still happily stumbling through Spanish, Telugu, and Malayalam. That slow, enjoyable business of actually learning a language convinced me practice has to be something you want to do, which is what we're building at Lingoken. I write here about making language activities people genuinely enjoy.
Build your first game free
Turn your next lesson into a game your students actually want to play. No account needed to try it, and no student logins.
Pick a template