Screens generate.
Paper teaches.

The repository ships a /langwich slash command. Claude asks for your languages and level once, suggests a few story ideas, writes the JSON, validates and repairs it, looks over the language and the facts, and renders the PDF — all in one conversation. This is the most comfortable way in.

git clone https://github.com/joernmht/langwich
cd langwich
python3 -m venv .venv
.venv/bin/pip install -e .
claude
> /langwich

Add a topic, a photo path or “next episode” after /langwich. Claude uses the langwich in .venv (and installs it there if it is missing). Your defaults are kept in .langwich/profile.json. Works in the Claude Code CLI, the desktop app and on claude.ai/code.

Any chat assistant that follows long instructions can write a worksheet — ChatGPT, Gemini, Claude.ai, Le Chat. langwich prints the brief; you carry it over and bring the JSON back.

1. Print the brief with your languages, level and topic.
2. Paste prompt.txt into a new chat, or attach the file. Starting from a photo? Add --image — langwich copies the photo to data/pictures/ — attach the photo as well, and use a model that can see images.
3. Save the reply as data/tides_en_de.json. Just the JSON object is best; if the model wrapped it in a code fence or chatted around it, langwich reads the object anyway and warns.
4. Check it. If langwich finds problems, paste the repair prompt into the same chat and save the new reply over the file.
5. Typeset it.

# 1. the brief (with a photo: add --image harbour.jpg and attach harbour.jpg too;
#    save the JSON in data/, where the copied photo is)
langwich prompt --source en --target de --level B1 --topic tides -o prompt.txt

# 4. check the JSON the AI wrote; if there are problems, make a repair prompt
langwich validate data/tides_en_de.json
langwich validate data/tides_en_de.json --prompt > repair.txt

# 5. typeset it → data/tides_en_de.pdf
langwich render data/tides_en_de.json

Add --save-profile once, and langwich remembers your languages, level, colour choice and device (never the story frame): next time langwich prompt --topic volcanoes is enough.

No cloud at all: langwich works just as well with a model on your own computer — in Ollama, LM Studio or llama.cpp. Small models do better with the shorter brief, and langwich’s repair loop catches what they get wrong.

# once: a copy of the model with a 32k context (Ollama's default is far too small)
printf 'FROM gemma3:12b\nPARAMETER num_ctx 32768\n' > Modelfile
ollama create gemma3-32k -f Modelfile

# a brief half as long, for models of 7–14B parameters
langwich prompt --compact --target es --level A2 --topic "a night market" -o prompt.txt

# paste prompt.txt into the model's chat, save its answer, then check it
langwich validate data/night_market_en_es.json

# problems? paste this repair prompt into the same chat, save the reply over the file
langwich validate data/night_market_en_es.json --prompt

langwich render data/night_market_en_es.json

• Context: the compact brief is about 3,500 tokens and a finished worksheet another 6,000–12,000, so give the model at least 16k tokens of context — 32k for the repair loop, which sends the whole JSON back. A cut-off answer is not valid JSON. In Ollama, use the 32k copy made above (or start the server with OLLAMA_CONTEXT_LENGTH=32768 ollama serve); in LM Studio, set the context length when you load the model.
• Bigger is better: larger models tell better stories and can take the full brief — leave out --compact.
• Photos need a vision model, such as Gemma 3 or Qwen2.5-VL.

A fresh worksheet on your desk every morning. Claude Code’s print mode runs the whole loop unattended — it only needs permission to run langwich and to write into data/:

#!/bin/sh
# ~/langwich/daily.sh: one new worksheet a day, saved as data/<date>.pdf
cd "$HOME/langwich" || exit 1
PATH="$HOME/langwich/.venv/bin:$PATH"   # the langwich installed in .venv
day=$(date +%F)
langwich prompt --target de --level B1 -o "data/$day-prompt.md"
task="Read data/$day-prompt.md and write the worksheet it asks for to data/$day.json.
Choose a topic that no other worksheet in data/ has used yet.
Then run 'langwich validate data/$day.json' and fix what it reports until there are no errors.
Finally run 'langwich render data/$day.json'."
claude -p "$task" --allowedTools "Bash(langwich:*)" "Edit(data/**)"

Make it executable (chmod +x ~/langwich/daily.sh) and add two lines with crontab -e. Cron starts with a bare PATH, so name the folder that holds claude (command -v claude shows it); the script adds langwich’s .venv itself:

PATH=/usr/local/bin:/usr/bin:/bin:/home/you/.local/bin
30 6 * * * $HOME/langwich/daily.sh >> $HOME/langwich/data/cron.log 2>&1

The permissions are deliberately narrow: Bash(langwich:*) lets Claude run langwich and nothing else, Edit(data/**) lets it write only inside data/; reading files needs no permission. Log in to Claude Code once as the user the cron job runs as.

Rather stay offline? The same loop with a local model through Ollama — up to three repair rounds, then render:

#!/bin/sh
# ~/langwich/local.sh: the same with a local model (Ollama), fully offline
cd "$HOME/langwich" || exit 1
PATH="$HOME/langwich/.venv/bin:$PATH"
model=gemma3-32k   # gemma3:12b with a 32k context, made once:
ollama show "$model" >/dev/null 2>&1 || {
  printf 'FROM gemma3:12b\nPARAMETER num_ctx 32768\n' > data/Modelfile
  ollama create "$model" -f data/Modelfile
}
day=$(date +%F)
langwich prompt --compact --target de --level B1 -o "data/$day-prompt.md"
ollama run "$model" --format json < "data/$day-prompt.md" > "data/$day.json"
for round in 1 2 3; do
  langwich validate "data/$day.json" && break
  langwich validate "data/$day.json" --prompt > "data/$day-repair.md"
  [ -s "data/$day-repair.md" ] || break   # nothing the model can repair
  ollama run "$model" --format json < "data/$day-repair.md" > "data/$day-fixed.json" || break
  mv "data/$day-fixed.json" "data/$day.json"
done
langwich render "data/$day.json"

Any model you have pulled works in place of gemma3:12b — keep the 32k context (PARAMETER num_ctx 32768): the repair round sends the whole worksheet back. ollama run … --format json keeps the answer to pure JSON, and the loop stops when there is nothing the model can repair. Pair either script with an e-paper tablet or a duplex printer, and the day starts with a worksheet instead of a feed.