How many German words do you actually need for B1? The numbers behind my vocabulary app

Nalivator3000 — Sun, 04 Oct 2026 13:13:40 +0000

I am a solo developer living in Germany, and in the evenings I build LexyBooster, a German vocabulary app with spaced repetition (https://lexybooster.com). Before writing any code for the word lists I had to answer one question for myself, because every expat I know asks it sooner or later: how many German words do you need for B1, and which ones first?

Here is what I found, and how I turned it into data.

The short answer: about 2,400 to 2,500 words, but not all at once

The Goethe-Institut B1 word list has roughly 2,400 entries, and most course books land around 2,500 words for B1. That number scares people, but it hides two things:

a large part of it is small, high-frequency words (articles, pronouns, prepositions, modal verbs) that you pick up in the first weeks anyway;
the exam does not test the list word by word. It tests whether you can handle everyday situations: forms, phone calls, appointments, short letters.

So the useful question is not "how do I memorize 2,400 words" but "which few hundred words unlock the situations I face every week".

I wrote the long version with the breakdown by level here: https://lexybooster.com/how-many-german-words-for-b1.html

Start with the situations, not the alphabet

When I looked at what newcomers in Germany actually struggle with, the same scenes came up again and again. Each one needs a surprisingly small vocabulary.

Registering your address (Anmeldung). About 30 words get you through the Bürgeramt: Termin, Meldebescheinigung, Wohnungsgeberbestätigung, Einzugsdatum, Reisepass, Unterschrift, Hauptwohnsitz. If you know these, the form stops being a wall of text. My list for this scene: https://lexybooster.com/german-words-for-anmeldung.html

Doctor visits and health insurance. Krankenkasse, Versichertenkarte, Hausarzt, Überweisung, Rezept, Krankschreibung. Most first appointments fail not on grammar but on these six words.

Work and the job center. Arbeitsvertrag, Probezeit, Kündigung, Gehaltsabrechnung, Steuer-ID. You will see them in your first month whether you like it or not.

I grouped the first 100 words an expat really needs into one page: https://lexybooster.com/essential-german-words-for-expats.html

How the app turns this into a schedule

From the developer side, the interesting part is that the word lists are just data, and the schedule is just math:

Words live in thematic collections (everyday situations, work, health) with a CEFR level.
New words come in small batches from the collection you choose, not in alphabetical order.
Reviews follow a spaced repetition schedule: a word you got right comes back after longer and longer gaps; a word you missed comes back the same day.
Every word has audio, because half of the problem at the Bürgeramt is recognizing a word you can read but cannot hear.

With 10 to 15 minutes a day and about 10 new words a day, the core scenes take a few weeks. The full B1 list takes months, and that is fine: by then you are already using German at the office and at the doctor, which is the point.

What I would tell a friend who just moved here

Do not start with the full B1 list. Start with the scene you will face next week.
Learn words with their article and plural (der Termin, die Termine). It costs nothing extra and saves you later.
Listen to every word at least once. Reading-only vocabulary disappears in a real conversation.
Review a little every day instead of a lot on Sunday. Spaced repetition only works if you show up.

If you are learning German in Germany, I would love to hear which situation was hardest for you in the first months. And if you build learning tools yourself: what do you use for scheduling reviews?

I run a one-person "dark factory" of AI agents for $0.015 per task. Here is what broke.

Nalivator3000 — Fri, 02 Oct 2026 09:21:24 +0000

I am a solo developer in Germany. In the evenings I build LexyBooster, a German vocabulary app with spaced repetition (https://lexybooster.com). A month ago I stopped writing most of the code myself and set up a small "dark factory": a pipeline of AI agents that picks up GitHub issues, writes the code, runs the tests, reviews the pull request and merges it when every gate is green.

The setup is boring on purpose:

one Hetzner VPS (4 vCPU, 8 GB RAM, about 12.60 USD a month);
Archon runs the workflows, Paperclip keeps budgets and schedules, LiteLLM routes every call to the cheapest model that can do the job;
free models first (Groq, Cerebras), Gemini Flash-Lite when they are not enough, and a Claude subscription for planning and review;
every merge decision is made by a script, never by a model: approve from the strong reviewer, green checks, no deleted definitions, no forbidden paths.

What the numbers look like after a month:

a typical task (one or two files plus a test) costs about 0.015 USD in API calls; the cheapest ones cost 0.007;
the biggest saving was not a better model but smaller tasks: when an epic is cut into tasks of at most four code files, a single cheap call writes a working patch in 40 to 50 percent of tasks;
the factory built six epics of a multiplayer Godot game mostly without me, plus SEO pages for the app.

What broke, and how I fixed the tool instead of the task:

The prepaid model account ran dry and nobody noticed. For two days every paid call failed, the pipeline silently fell back to the subscription model and kept merging. Fix: a monitor check that reads the gateway logs for "credits depleted" and pings me directly.
A PR waited two days for a human who never got the message. The alert went into the inbox of an "agent that fixes alerts", and that agent is not allowed to merge config changes. Fix: alerts that need money or a merge decision skip the agent and go straight to me.
The improver agent kept improving the same module. It saw its own commit from last night and proposed a follow-up, three nights in a row. Fix: a 14-day cooldown per module.
Cheap agents looping on hard tasks. When the one-call patch failed, a cheap agent kept iterating for up to 33 calls and still often needed the stronger model. Fix: after the one-call patch and two quick repairs, the subscription model takes over; the cheap agent runs only when the subscription window is closed.

If you are building something similar, the one lesson I would keep: let models write, but let scripts decide.