I have been interested in languages for most of my life, and not casually. Spanish and French are the two I have pursued most seriously, but over the years I have dabbled in at least half a dozen others, and that is before you count my occasional fascination with constructed languages, or conlangs.1 I am the kind of person who buys language courses, studies grammar, learns vocabulary, listens to lessons in the car, reads about how languages work, experiments with different learning systems, and every so often decides that this time I am really going to become fluent. I genuinely enjoy all of it.
And yet, after all of that study, I have spent years running into the same wall. I do not get enough opportunities to actually speak.
That sounds almost embarrassingly obvious when I write it down. You cannot learn to have conversations without having conversations. But for an adult learner who does not live in a place where the target language surrounds them, that is a much bigger problem than it sounds. I can open a book whenever I want, listen to audio lessons whenever I want, and study vocabulary at six in the morning or practice verb conjugations before bed. What I cannot easily do is summon a patient French speaker at 7:15 on a Tuesday morning because I happen to have twenty minutes and feel like talking about my weekend. That has always been the missing piece.
I Have Learned a Lot of Language
The frustrating thing is that I am not starting from zero. I know a fair amount of Spanish and a fair amount of French, and I have wandered into other languages often enough to recognize the patterns, the similarities, and the recurring problems that come with trying to make a new language feel natural. Spanish and French are where the frustration is most obvious, because I know enough of both to see how much I know and how much of it disappears the moment I have to use it in a real conversation.
Depending on the subject, I can read things that would surprise someone who had only heard me struggle to speak. I understand grammar, I recognize a large vocabulary, and I can usually understand far more than I can comfortably produce. Give me time to think about a sentence and I can figure out what I want to say. Then someone actually talks to me, and suddenly all of that knowledge feels locked in a filing cabinet somewhere in the back of my brain. I know the word. I know that I know the word. I cannot find the word. Or I start a sentence confidently and realize halfway through that I have no idea how to finish it. Or someone answers me in perfectly ordinary language at perfectly ordinary native speed, and my brain catches three words out of eleven.
Then comes the awkward little moment. "Uh." "Comment dit-on." "¿Cómo se dice?" Then English, then frustration, and then the conversation stops being a conversation.
This has happened enough times that I have come to believe there is a real distinction between knowing a language and being able to use one. The two overlap, obviously, but they are not the same skill. I have spent years building knowledge. What I lacked was time developing reflexes.
Traditional Language Learning Has Helped Me
I do not think the tools I used failed me. Quite the opposite. Pimsleur helped me enormously, because audio courses forced me to retrieve language instead of simply recognizing it, and they built pronunciation, sentence patterns, listening, and the ability to produce a response without staring at a page. Books helped, vocabulary study helped, grammar helped, and reading helped. Experimenting with other languages helped too, because every new language gives you another angle on how people construct meaning, what languages share, and the strange ways they differ. Every piece added something.
The problem is that none of it completely solved the problem I actually had. Those were all parts of the puzzle, and I was still missing immersion. Not the romantic version where I move to Madrid for six months and drink coffee in a neighborhood café every morning, but something much simpler. I needed to talk, a lot. I needed hundreds of ordinary conversations. I needed to talk when I was tired and when I could not remember a word, to tell the same kinds of stories over and over until the language stopped feeling constructed, and to do it with someone who would let me stumble through a sentence without turning the moment into a grammar lesson. Most of all, I needed it available often enough that speaking could become normal. That has always been the hard part to arrange.
The Awkward Part of Learning a Language
One thing language courses cannot fully simulate is how uncomfortable real conversation is. You feel much less intelligent in a language you are still learning, and that may be one of the hardest parts. In English, I can explain a complicated technical idea, argue about politics, tell a story, make a joke, shift my tone, or find exactly the word I want. Put me in a French or Spanish conversation and I am suddenly trying to explain that same thing with the linguistic sophistication of a six-year-old. The thought is not simple. The language available to me in that moment is.
That gap is genuinely frustrating. Sometimes I know exactly what I want to say and do not have the machinery to say it. Sometimes I have the machinery but cannot retrieve it fast enough. Sometimes I start talking and discover halfway through that I have built myself into a grammatical corner, so I have to back out, start again, simplify, reach for the wrong word, talk my way around the thing I actually mean, and occasionally surrender and say it in English. That is not a failure of language learning. I increasingly think that is language learning. You have to spend time inside that uncomfortable space, and the real problem is finding enough chances to do it.
I Did Not Want Another Tutor
That frustration is what led me to start experimenting with conversational GPTs for language learning. My first instinct was the obvious one: build an AI language tutor. But the more I thought about it, the less interested I was in a tutor. I already have lessons. I can ask ChatGPT to explain the subjunctive whenever I want, generate vocabulary lists, pull conjugation tables, or produce exercises. All of that is easy. What I wanted was someone to talk to.
That distinction shaped everything as I started writing the instructions for the GPTs. I did not want every mistake corrected, I did not want every exchange to end with a question, and I definitely did not want "Excellent! Great job! Now let's practice another sentence!" Nobody talks like that. I wanted a conversation, which means sometimes the AI asks a question, sometimes it just reacts, and sometimes it laughs, disagrees, offers an opinion, or stays with something interesting I said. I wanted to talk about software one day and my granddaughter the next, without the conversation lurching into a new subject because the lesson plan decided it was time to practice restaurant vocabulary.
The learning needed to happen inside the conversation instead of constantly replacing it. That became one of the central ideas behind the prompts I eventually wrote. Conversation is the product. The learning is still there. It just does not always announce itself.
Teach Me Without Constantly Teaching Me
This is one of the things I thought about most while designing the prompts. I absolutely want the GPT to teach me. I just do not want it to behave like a teacher every thirty seconds, because there is a difference between the two.
Suppose I say something wrong in Spanish. A traditional instructional exchange would stop me, name the error, explain the rule, give the correct form, and ask me to repeat it. Sometimes that is exactly what I need, but not every time. In a real conversation, it is often more useful for the other person to simply answer using the correct form. I hear it, I see how it fits, and the conversation keeps moving. Maybe the same structure comes back five minutes later, maybe I try it myself, maybe I get it wrong again, and maybe this time I get it right.
That is the learning experience I have been trying to build. Not the absence of teaching, but teaching woven into the conversation so naturally that I sometimes forget it is happening.
Do Not Correct Everything
This was another deliberate choice. I make a lot of mistakes when I speak another language, and if every mistake becomes an event, the conversation gets exhausting. Imagine talking in English with someone who interrupts you every twenty seconds with "wrong tense," "wrong preposition," "you mispronounced that," "you should have used a different article." Technically useful, socially unbearable, and eventually you just stop talking.
So I wanted the GPTs to make a judgment call. Did the mistake change what I meant? Did it make me hard to understand? Do I keep making the same one? Is it useful enough that correcting it right now would actually help? If so, maybe correct it. Otherwise, keep talking. The goal, for me, is not perfect sentences. The first goal is to become comfortable producing sentences at all, and accuracy can improve inside that process.
Please Do Not Interview Me
Another thing I learned quickly is that AI loves questions. Tell a GPT to be conversational and you often get a rapid-fire string of them: what did you do today, what did you eat, did you enjoy it, who did you go with, what are you doing tomorrow. That is not a conversation. That is an interrogation with good manners.
Real conversations breathe. Sometimes someone asks a question, sometimes they tell a story, sometimes they make an observation or say something funny, and sometimes they just react and leave enough silence for the other person to decide where things go next. That mattered enough that I built it directly into the prompts: do not ask a question after every response, do not constantly change topics, stay with interesting things, remember what I told you, and talk to me like someone who was actually listening. These sound like small design decisions. For me they are the difference between doing a language exercise and having a conversation.
Meet Me Where I Am Without Announcing It
I also did not want a placement test. I have taken enough tests, I know roughly where my skills are, and the honest answer depends on what you are measuring anyway. My reading ability is not my speaking ability, my vocabulary is not my retrieval speed, and my ability to follow a carefully spoken recording is not my ability to follow someone talking naturally.
So I wanted the GPT to figure it out quietly. Start talking, see what I understand, notice how I answer, simplify when I am struggling, and stretch me a little when I am comfortable, introducing more natural language over time. Do not announce, "Based on your responses, I estimate that you are B1." Just talk to me. A good human conversational partner does this almost without thinking, constantly adjusting to the person across from them, and I wanted the AI to do the same.
Let Me Struggle a Little
This one is tricky. I do not want the AI to rescue me too quickly, because struggling to find the language is part of the exercise. At the same time, I do not want to spend two minutes stuck because I cannot remember the word for "screwdriver." Sometimes I genuinely need help.
So one of the ideas built into these GPTs is a kind of graceful recovery. Try again in simpler Spanish or French, give me another way to understand it, use an example, and when that is clearly not working, use a little English and then go back. English is not forbidden. It is a bridge. Rigid immersion can become absurd: if I spend ninety seconds misunderstanding a sentence that six English words would clear up, we are not learning efficiently. But switching to English the instant I hesitate defeats the whole purpose. The interesting part is finding the space between those two extremes.
Remember the Conversation
This may be one of the most underrated parts of immersion. Real relationships have continuity. If I told you yesterday that my granddaughter was going to a trampoline park, you might ask later whether she had fun. If I mention that I am preparing for a hard meeting at work, you might remember it the next time we talk. Even inside a single conversation it matters, because nothing makes an AI feel more artificial than telling it something important and having it ask for the same thing ten minutes later.
So I wanted these conversational partners to pay attention: remember the people I mention, remember my interests, remember what we were talking about, and use those things naturally. Not because memory itself teaches Spanish or French, but because meaningful conversation does, and meaningful conversations are connected.
I Think AI Changes Something Here
I am careful about claiming AI has "solved" anything, and it has not solved language learning. You still have to learn vocabulary, listen, read, wrestle with grammar, and speak, especially speak. But something feels different now. For the first time, I have a patient conversational partner available almost whenever I want one. I can speak French at six in the morning and practice Spanish while driving, and if tomorrow I decide to revive one of the half-dozen other languages I have dabbled in, I can do the same thing there.
I can stop and ask what something means. I can make the same mistake ten times without embarrassing myself. I can ask for more correction one day and less the next, and I can talk about subjects I actually care about instead of whatever happens to appear in Chapter 7. And I can do it every day. That availability matters more than it sounds. For someone who has spent years fascinated by languages but never able to find enough consistent immersion, it feels genuinely new.
Not because AI is better than talking to real people. It is not, and real human conversation has things AI cannot replace. But that was never my problem. My problem was that I did not have enough of those conversations. What AI finally gives me is something I have never had before: volume. Hundreds of conversations, potentially thousands. Enough repetition that retrieving a sentence might stop feeling like solving a puzzle. Enough exposure that ordinary spoken French starts to sound ordinary. Enough practice that Spanish stops feeling like something I know and starts feeling like something I can simply use. Enough mistakes that making them stops feeling like a big deal. Enough time in the language that I might eventually stop translating everything in my head before I speak.
That possibility did something I did not expect. It made me excited about language learning again. I find myself wanting to speak French, wanting to practice Spanish, wanting to dust off languages I have only dabbled in and see what happens when conversation is no longer the scarce resource, maybe even wanting to try entirely new ones. That renewed curiosity is not a small thing.
So I Built María and Amélie
I have been experimenting with two GPTs built around these ideas. Conversaciones con María2 is my Spanish conversational partner, and Parle avec Amélie3 is the French one. I hesitate to call either a finished language-learning system, because that is not really what they are. They are experiments, my attempt to take all the little frustrations I have collected over years of study and turn them into a different kind of experience: conversation first, teaching when it helps, correction without constant interruption, difficulty that adjusts quietly, English when it is genuinely needed but not as an automatic escape hatch, enough patience to let me search for a word, and enough personality that the whole thing feels less like an exercise and more like talking to someone.
Most importantly, they are available. When I have twenty minutes, there is someone to talk to. That may sound minor. After years of trying to learn languages without enough chances for immersion, it does not feel minor at all.
I am also keeping the prompts and related experiments on GitHub,4 partly because I expect them to keep changing. The interesting question is not whether I have found the perfect prompt, because I obviously have not. It is how much better this can get as I keep noticing what makes a conversation feel natural, what helps me learn without breaking the flow, and what gets in the way.
That is really what these prompts are. Not my attempt to define the correct way to learn a language, but a list of things that turned out to matter to me after years of trying. Talk to me. Be patient. Do not turn every mistake into a lesson. Teach me without constantly reminding me that you are teaching me. Let me struggle, but help me when I am stuck. Remember what I say. Stay with interesting conversations. Give me enough chances to speak that speaking eventually becomes normal.
For years, immersion was the one part of language learning I could never manufacture for myself. For the first time, I think maybe I can. And honestly, that makes me want to start all over again.
Frequently asked questions
Does AI replace traditional language learning?
- No. You still have to learn vocabulary, listen, read, and struggle with grammar, and courses like Pimsleur, books, and vocabulary study all do real work. AI fills a different gap: it gives an adult learner enough low-stakes conversation practice that speaking can finally become normal. It supplements the study; it does not replace it.
What is the difference between knowing a language and being able to use one?
- Knowing a language is recognition: understanding grammar, reading, and recognizing a large vocabulary. Using one conversationally is retrieval under time pressure, producing the right words fast enough to keep a real conversation moving. The two overlap but are not the same skill, and most classroom study builds knowledge far faster than it builds reflexes.
Why build a conversational partner instead of an AI tutor?
- Lessons are already easy to get: any capable model can explain the subjunctive, generate vocabulary lists, or produce exercises on demand. The scarce resource for an adult learner is conversation itself. A partner whose whole job is to talk with you, with teaching woven in rather than announced, targets the part that traditional tools cannot easily provide.
How should an AI conversation partner handle mistakes?
- Selectively, not constantly. Correct a mistake when it changes your meaning, makes you hard to understand, keeps recurring, or is genuinely useful to fix right now. Otherwise, model the correct form in the reply and keep the conversation moving. Interrupting every twenty seconds is technically useful and socially unbearable, and it eventually makes people stop talking.
Should the AI use English while practicing another language?
- English works best as a bridge, not a forbidden word or an automatic escape hatch. When you are stuck, the partner should first try simpler phrasing, another explanation, or an example, and only fall back to a little English when nothing else is working, then return to the target language. Rigid immersion wastes time; instant English defeats the purpose.
What does AI actually add to language learning?
- Volume. A patient conversational partner is available whenever you have twenty minutes, which means hundreds or potentially thousands of ordinary conversations. Enough repetition and exposure that retrieving a sentence stops feeling like solving a puzzle, ordinary spoken speech starts to sound ordinary, and making mistakes stops feeling significant.
Footnotes
- Constructed languages (conlangs) invented languages such as Esperanto, Klingon, or Tolkien's Elvish. ↩
- Conversaciones con María the Spanish conversational partner GPT. ↩
- Parle avec Amélie the French conversational partner GPT. ↩
- prompts and related experiments on GitHub ↩