language learning in public w/ jack

Learning languages in public with current focus on Brazilian Portuguese. Here is the setup. Background: I speak English, French, German, Chinese, Italian and Norwegian to varying levels. Portuguese is next. Approach: Claude in Cowork acts as a persistent second brain across every session. I have been using Claude heavily for work and found it to be the best tool for reading and writing to files in a way that creates a self-updating memory, the same way a human learns and updates their own. It reads `CLAUDE.md` before we start, logs every sentence I learn in `Review_Queue.md`, tracks recurring errors, and keeps a detailed `Session_Log.md`. Nothing resets between sessions. More on why this is different below. Goal: advanced conversational PT for Brazil trip by September 2026. About 2-5 hours a week. What I'll post here: honest updates on what is working, what is not, and how the system evolves. No hype, no hiding the bad sessions.

26 May 2026 | system update After today's tutor session I generated a NotebookLM prompt directly from the sentence queue. The source is 12 sentences from the caderno, plus the patterns that mattered: gender agreement, masculine nouns ending in "a", past tense narration, and "inscrever se em". The prompt tells NotebookLM to go sentence by sentence and explain what is interesting. This is the part I like. The tutor gives me messy real material. Codex turns it into `NotebookLM_Prompt_26_May_2026.md`. NotebookLM turns that into audio review I can listen to before the next session.

26 May 2026 | self-study Did a review test today using sentences from earlier Portuguese sessions. This is where having AI track everything I am doing and write to my folder structure becomes useful - all prior learnings and issues are logged and can easily be brought back in review session. In terms of progress - big grammar topics are ok. Small production mechanics are not automatic yet. I mixed up "moro" and "morro" 🤣, used "pra" where I needed "por", dropped a repeated "não", and reached for "dizerem" instead of "disseram" 😬 ... So `Review_Queue.md` now has some rows marked reviewed and others marked needs review. The next focus will be past tense irregulars, verb prepositions, and fixed chunks (e.g. "na maior parte das vezes") until they come out cleanly.

26 May 2026 | tutor session Second tutor session had a very ordinary but useful pattern: gender agreement kept surfacing in sentences I actually need. "Meus pais já são aposentados." "A família dela é católica." "Minha namorada também é ateia, eu sou ateu." "Eu sou viciado em podcasts." None of this is conceptually hard. The work is producing it cleanly while talking. Logged 12 new sentences in `Review_Queue.md` and created a new NotebookLM prompt for the batch.

@punkerlabrat folder in google drive is the key - can acccess it from any (current or future) AI tool that has desktop access

@fluency_pending The shared folder part is huge. Half the pain is just dragging context back into the room every time.

@jegaevi take it as a blessing that you dont need to learn as many pinyins - and just forget the tones nobody will care that much (controversial but pareto) 😹

Chinese stop confusing me with words that are almost the same challange

23 May 2026 | system update After the first tutor session I realised the system had no pipeline for what happens after a class. The caderno comes in, sentences get logged, and then nothing forces them back in front of me before the next session. So now `CLAUDE.md` has a six step caderno dump workflow: extract sentences, update the verb log, check for recurring errors, log homework, generate a NotebookLM export, write the session log. There is also a pre-session prep template that pulls unmastered sentences and recurring errors into a briefing before each class. The first NotebookLM export is ready. NotebookLM (notebooklm.google) lets you paste in source material and it generates a podcast style conversation between two AI hosts who discuss it. So I paste in twenty annotated sentences from my tutor session, and it produces a 5 to 10 minute audio episode where the hosts read each sentence in Portuguese, explain the grammar, and draw connections between them. Passive review I can listen to on a walk before next week's class.

22 May 2026 | tutor session First tutor session. My tutor keeps a "caderno," a shared Google Doc where she writes down corrections every time I try to produce a sentence. It is a running record of what I got wrong and what the right version looks like. What surprised me: the tutor introduced subjunctive and conditional structures in session one. "Não é fácil encontrar um professor que eu goste." These are supposedly intermediate or advanced grammar. But if you already have French, the patterns are immediately recognisable. After the session I fed the caderno into Claude. It extracted 20 sentences, updated the verb log (10 of 16 priority verbs now encountered), and logged seven new grammar patterns. The tutor does not know about the system. She just hands me a document.

kinda, the mapping gets messy because it's going word-by-word. The whole chunk 我在飯店對街家店裡看到的 is one "adjectival unit", it'd be clearer to colour it as a single block and map it to "that I saw in a shop across the street from the hotel" on the English side. also… are you using traditional characters?? 😵

@fluency_pending yes, I think there is an error in the results generated by ChatGPT. How about this one?

ga jadi guys, aku minta chat gpt buatin sentence maps english to mandarin 🥰. Semoga bermanfaat ya teman-teman 😆, mungkin yang lain boleh minta tolong koreksi bener ngga yang dibuat chat gpt untuk kalimat upper-intermidiate sm complex #studytwt #langtwt #ElbyAndBooks #mandarin

good list, I'd add a 5th: use a mem (mnemonic anchor). some people visualise the character's radical + components and what they literally mean. Others anchor to the pinyin sound. E.g. 你 (nǐ) = "you" → kick YOU in the NI (knee). this helped me learn 3,000+ characters when @ uni

How to never forget a Chinese word: 1. Learn it in a sentence, not isolation 2. Use it in conversation within 24 hours 3. Connect it to a personal memory 4. Hear it from 3+ different people Words without context die. Words with stories live forever.

@xiaonixia your notes are very aesthetic, well done, you have so much patience haha

learning how to say 'again' with 又 and 再🧩 #studytwt

@profhistorygeek flash card app of choice = ?

Latin study update: I continue to use Duolingo daily and have started with Lingua Latina. I also caved and purchased a copy of Wheelock in kindle format to read on the commute to work. I have started a vocab flash card deck too.

@lingdiscovery @Starkey_Comics this is incredible haha, i love these charts cc: @Starkey_Comics

All the ways that the Proto-Indo-European word for ‘one’ surfaces in Modern English, in collaboration with @Starkey_Comics !

for anyone finding Mandarin word order tricky, a few things that helped me: - basic order is S+V+O, same as English: 我吃苹果 = I eat an apple - time & place go before the verb: 我昨天在家吃饭, not 我吃饭昨天在家 - "manner adverbs" also before the verb: 慢慢地吃 (slowly eat), not 吃慢慢地 - relative clauses flip and go before the noun: 我看的书 = "the book [that] I read" think the golden rule: almost everything that English puts after the verb, Mandarin puts before it ↔️

@jellykeikeu @bangsawanetijen iya kak, aku juga selalu bingung ngerangkai kalimat mandarin tuh gimana 😭, trus baru inget ada orang pernah bikin sentence maps kek gini deh

@danialnamkr boss level!!

receipts 👇 🗣️ swain 1985 (output hypothesis): en.wikipedia.org/wiki/Output_hy… 👁️ schmidt 1990 (noticing hypothesis): doi.org/10.1093/applin… ✏️ lyster & saito 2010 (corrective feedback meta-analysis): doi.org/10.1017/S02722… if you only have time for one, lyster & saito 2010. 15 studies meta-analyzed. effect size d=0.74 (medium-large), meaning learners getting corrective feedback outperformed controls by ~3/4 of a standard deviation.

SLA (second language acquisition) research has been clear for 40 years. acquiring a language requires 3 things, not 1 1. 🗣️ output: saying it out loud (swain 1985) 2.👁️ noticing: catching your own errors (schmidt 1990) 3. ✏️corrective feedback: fixing them on the spot (lyster & saito 2010) most apps only cover output. you speak, you move on. your brain keeps encoding broken patterns. 10 min of all 3 beats 1 hour of just speaking. (this is exactly why i built langtwo) #langtwt #studytwt

@lingdiscovery love it, amazing -- will inspect your Following then as assuming it's a goldmine

Thank you! I just have a list of sources and alerts I check daily. Plus I just come across a lot of stuff randomly in my own social media feeds. Then I have a Notion database where each item gets its own entry. At the end of the week I just compile the entries into a digest. I've got a workflow that I check off for each item, so it's pretty streamlined at this point.

🧬 New evidence that language evolved before humans did 🤬 Universals of swearing across languages 🏛️ Does learning Latin improve your SAT scores? ✍🏼 What killed penmanship? 🧩 Try your hand at puzzles from North America’s biggest linguistics competition

@yubinsann and your pronunciation of umlauts ä ü etc is very goood

@yubinsann great work and good pronunciation! One pointer - the “z” sound could be pronounced a bit as “ts” in English So, Zentrum becomes “tsen-trum” Keep going!!

I’ve decided to start sharing my German speaking practice here on #langtwt! 🇩🇪 ​My pronunciation isn't that good yet, but I’m excited to look back and see the improvement over time. Constructive feedback is welcome, but go easy on me - it’s my first time recording #studytwt

@PolyglotZone please report back!!

I am considering getting LingQ again for Mandarin reading. Does anyone like the current version for Mandarin? I hear the dictionary is context sensitive now. That's a minimum requirement for me with Mandarin readers now. None of the "click one character, get 100 meanings" stuff