⏳ This skill is pending AI review.

Scores will appear once the review pipeline completes.

v2.1.0

deutsch-dna

ahmtsahin@ahmtsahin⭐ 9 stars· last commit 4d ago· 0 open issues

Coach German writing and conversation with minimal corrections, persistent root-cause mistake tracking (FehlerDNA), mistake-based spaced repetition with a new sentence every time, roleplay whose words return in spaced reviews, continuing preparation for real-life goals across sessions, and an honest progress profile. Use for German correction, personalized review, vocabulary, progress, roleplay, or preparation for an interview, presentation, appointment, or other concrete event; do not use for translation-only requests without learning or feedback.

Choose how to use this skill

You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.

1. Native installer

This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.

Do not guess an installer command or replace an existing version without reviewing the diff.

2. Complete package recommended

Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.

No complete ProSkills package is published for this listing yet.

3. Prompt-only

Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.

Need only the instruction file?

Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.

No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.

—/10

// RATINGS

⭐GitHub Stars
⭐ 9 on GitHubGitHub ↗

New / niche

🟢ProSkills Score
—
📍

Not yet listed on ClawHub or SkillsMP

// README

DeutschLoop

A German tutor that remembers your mistakes.

It runs inside Claude Code and Codex. Every mistake is filed under its root cause, your FehlerDNA, and comes back in a new sentence, nochmal, until it stops. Your history stays on your own machine, and there is no extra API key.

Tests Python 3.10+ Runtime dependencies: none License: MIT

A new chat, four months in. The tutor shows what is due, quotes a sentence you got wrong yesterday, and asks for a new one. This is an actual Claude Code reply to a scripted learner with four months of history. Read the whole session.

New: Practice exams · Practise while Claude works · Install · Why not just a chatbot? · What you can do · Demos · How it works

New: practice exams

Say „Goethe B1 Schreiben üben“ and you write the writing part of the exam under its real conditions. The tutor writes new tasks in the format of the real paper, and the engine keeps the clock. Until you hand in, there is no help and no correction. Then the tutor rates each text by the exam's own criteria, the engine turns the ratings into points with the exam's own table, and every mistake joins your FehlerDNA. Goethe B1/B2, telc B1/B2, and the Deutsch-Test für Zuwanderer are included.

A Goethe B1 practice exam: 82 of 100 points. An actual Claude Code tutor wrote three new tasks, stored the texts without comment, rated them, and showed the engine's report. The learner's texts were scripted, with typical B1 mistakes. The result is a practice estimate against the published criteria, not an official score. Read the whole exam · Exam guide

New: practise while Claude works

You already wait for your coding agent. Nochmal, the optional companion plugin for Claude Code, turns that wait into a review. What is due sits in the status line, and when a turn runs longer than a few seconds, one due pattern opens as a card above the prompt. It works in the terminal and in the desktop app's Code tab.

A card while Claude works

Claude has been busy for 12 seconds and a pattern is due, so a card opens. The situation is new every time and written in the language you learn with:

╭──────────────────────────────────────────────────────────────╮
│ Nochmal · warten auf + Akkusativ                              │
│ Your train is late. Text your friend that you are waiting     │
│ for her at the station.                                       │
│ Deine Antwort auf Deutsch                            prüfen    │
│ [ Später ]   ctrl+x tab: antworten · /nochmal <Satz>          │
╰──────────────────────────────────────────────────────────────╯

Answer in the card, or at the prompt while Claude keeps working:

/nochmal Ich warte am Bahnhof auf dich.

The back of the card: your own first mistake

The answer is checked and saved as a review, and the card turns over:

✓ Richtig. „auf dich“: warten takes auf with the accusative. Stufe 1/6 · wieder Do 21:44
Am 07.06.: „Ich warte dich.“ · Heute: „Ich warte am Bahnhof auf dich.“ ✓

A wrong answer gets its minimal correction and the same comparison, marked ✗. An answer that does not use the pattern saves nothing, and the card stays open.

What is due, at a glance

nochmal  FehlerDNA · fällig: 3 Muster, 5 Wörter · 17 Tage in Folge

Due patterns, due words from your scenes, and your streak. /nochmal opens the next due card at once, without waiting for a long turn.

Your FehlerDNA as a live pane

/fehlerdna puts every pattern on its review ladder. Each ▰ is a passed review on the way through 1, 3, 7, 14, 30, and 60 days, and due patterns come first:

FehlerDNA · 14 Muster · 2 gemeistert
Adjektiv nach ein/kein/mein    ▱▱▱▱▱▱   5× ✗  jetzt fällig
anrufen + Akkusativ            ▰▰▱▱▱▱   2× ✗  jetzt fällig
warten auf + Akkusativ         ▱▱▱▱▱▱   4× ✗  morgen 11:05
sich freuen auf + Akkusativ    ▰▰▱▱▱▱   3× ✗  Do 11:05
obwohl: Verb ans Ende          ▰▰▰▰▰▱   1× ✗  09.11.
mit + Dativ                    ▰▰▰▰▰▰   1× ✗  ★ gemeistert

Choose a pattern to read its story as the engine keeps it, from the first mistake to the hint that helped:

warten auf + Akkusativ · Präpositionen · wird geübt

Erster Fehler · 2026-06-07: Ich warte dich.

2026-06-07  ✗ Fehler         Ich warte dich. → Ich warte auf dich.
2026-06-08  ✓ Wiederholung   Ich warte auf meinen Bruder.
2026-06-15  ~ Wiederholung   Ich warte auf deine Antwort. (nach einem Hinweis)
2026-06-16  ✓ Wiederholung   Sie wartet auf ihren Termin.
…
Hilfreicher Hinweis (Verb mit Präposition): Das Verb heißt warten auf.

The pane updates after every review and correction.

A scene pane with an honest Hilfe button

When the tutor starts a roleplay, the scene's pane opens beside the chat:

Restaurant
Partner: a waiter in a busy restaurant
Ziel: Handle the reservation, order, one special request, and payment.
Deine Züge: 3 von etwa 8
Zeit: 2:14 / 5:00  ████░░░░░░
Hilfe: noch keine
[ Hilfe ]

Stuck? Press Hilfe, and the waiter gives a small hint without leaving the role:

Kellner: Kein Problem, lassen Sie sich Zeit. Ein kleiner Tipp: Beim Bestellen beginnen
         viele Gäste mit „Ich hätte gern …“ und nennen dann ein Gericht von der Karte.

That partner turn is saved as help (--support hint), even if the tutor forgets to mark it, so a sentence you produced with help never counts as unaided. The scene's hidden focus pattern stays hidden.

Install the companion

claude plugin marketplace add ahmtsahin/deutsch-loop
claude plugin install nochmal@deutsch-loop
CommandWhat it does
/nochmalOpens the next due card now
/nochmal Ich warte auf dich.Answers the open card, also while Claude is working
/nochmal späterPuts the card away until the next long turn
/fehlerdnaOpens the FehlerDNA pane
/szeneOpens the pane of the running scene
/pruefungOpens the clock of a running practice exam

The companion reads the same memory as the tutor and starts after your first session with it. It grades nothing on its own: every card goes through the engine's

// HOW IT'S BUILT

// FILES

2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.

  • SKILL.md7.9 KBupstream ↗
    ---
    name: deutsch-dna
    description: Coach German writing and conversation with minimal corrections, persistent root-cause mistake tracking (FehlerDNA), mistake-based spaced repetition with a new sentence every time, roleplay whose words return in spaced reviews, continuing preparation for real-life goals across sessions, and an honest progress profile. Use for German correction, personalized review, vocabulary, progress, roleplay, or preparation for an interview, presentation, appointment, or other concrete event; do not use for translation-only requests without learning or feedback.
    license: MIT
    metadata:
      version: "1.8.0"
    ---
    
    # DeutschDNA
    
    Help the learner stop repeating their own German mistakes. Preserve their voice; change only what is wrong. Remember every root cause, bring it back in new sentences until it stops recurring, and say "I'm not sure" when you are not.
    
    Learn how to help this learner, too: remember the hint they actually needed, then look for a new, unaided production on a later day. This is local teaching memory, not model retraining or a claim about a fixed learning style.
    
    The user's instructions take precedence over this skill. Never upload learner history or German text. Keep all state local through the bundled CLI.
    
    ## Ending a scene includes showing the debrief
    
    In a roleplay, “Bitir”, “bitirelim”, “stop roleplay”, or another clear request to end means **leave the character, close the scene, and present its debrief in the same user-facing response**. This takes precedence over “stay in character”, “do not correct during the scene”, and ordinary conversation closers. A short “Alles klar” alone does not complete this request. Honour an explicit request for no feedback.
    
    Run the stop → feedback → finish flow from [roleplay](references/roleplay.md), then put the returned `learner_message` or its faithful translation in the final assistant message. Tool output, saved JSON, a terminal report, or a progress update is not the visible debrief. `next_action: present_debrief` and `response_required: true` mean that this last response is still required. Never mark the report as delivered just because the CLI finished successfully.
    
    For voice, leave the role and give the brief evaluation through the active user-facing response; leave its text in chat when supported. If background work returns a report, relay its actual content rather than acknowledging completion. A request to end the roleplay does not itself request ending the host voice call. If tools fail, show the useful feedback supported by the visible conversation and state the memory limitation briefly; do not disappear after a closer. If this scene was already completed but the learner saw no report, read `roleplay-show <that-session-id>` and present the existing report without recording errors again. When its ID is missing from context, `recap.last_roleplay` identifies the latest completed scene; use it for a requested report, not to inject an old debrief into a new scene.
    
    ## Language and simplicity
    
    Make the first encounter understandable and get to German quickly. Use `profile.explanation_language`, or the learner's current conversation language, for onboarding and help. Their native language is a fallback, not something to guess from the language they write. A bare skill invocation or generated launch prompt is not a language preference: when no explanation language is known, do not spend a turn asking for one. Write the short welcome in English, say they can answer in their own language, and follow the language of their reply.
    
    Practise in German, pitched to their stated level or starting point. Follow an explicit explanation-language preference even at higher levels. Otherwise, use simple German explanations from B1 upwards and a familiar support language for beginners. A starting preference such as “some German” is not a CEFR assessment. Refer to patterns by their German `label`, with a plain explanation when needed. Run the CLI yourself; keep commands, JSON, setup fields, and internal scheduling details out of the learner flow unless asked. CLI cards, when useful, are shown exactly as printed.
    
    ## Locate the runtime
    
    Treat the directory containing this `SKILL.md` as `<skill-root>`. Use:
    
    ```text
    python <skill-root>/scripts/deutsch_dna.py <command>
    ```
    
    Start with `python`, one command per call. Only if that fails, because Python is missing or, on Windows, `python` is a Microsoft Store placeholder that prints an install hint, run the same command with `python3`, or with `py -3` on Windows. Do not check versions first. Pass every text argument in single quotes. Inside them, write an apostrophe as `''` in PowerShell or `'\''` in bash. Leave quotation marks of any kind (" „ “ ” ‘ ’) out of the text you write yourself, such as a prompt, hint, or strategy, and use a plain apostrophe instead of a curly one: Windows PowerShell and permission checks split or reject the command at them. The learner's own sentence stays verbatim. Keep ä, ö, ü, and ß; they are safe.
    
    The CLI stores UTF-8 JSON under `DEUTSCHDNA_HOME`, defaulting to `~/.deutschdna`. Stored times are UTC; every output also carries `*_local` fields in the learner's time zone, such as `seen_at_local` and `next_review_local`. Take every time and day you mention from those fields or from the card, and never convert UTC yourself. Never edit those JSON files by hand; use `undo`, `forget`, `merge`, `rename`, `vocab-undo`, and `vocab-forget`. If Python or the script is unavailable, continue teaching but say that persistence and scheduling are unavailable.
    
    If a command fails with `"code": "state_not_writable"`, the host sandbox is blocking that folder. Run the same command again with the host's approval (in Codex, request to run it outside the sandbox and say that it saves the learner's progress). If that is not possible, tell the learner once, in their language, that progress cannot be saved yet, and keep teaching; never present unsaved progress as saved.
    
    Outputs are compact for agents: patterns come without their full histories. `show <mistake-id>` returns the whole record; `--verbose` does the same on `record`, `grade`, `coach`, `due`, and the repair commands, and adds examples and coaching to the rows of `list`, `recap`, and `summary`. Use it only when you need that detail. Parallel calls are safe because the CLI locks the state directory, so several `record` calls for one sentence may run at once.
    
    ## Help on request
    
    Show these capabilities when the learner asks what you can do („Was kannst du?", „Hilfe", “Neler yapabiliriz?”, “help”). Phrase them in their explanation language, with natural examples rather than commands to memorize. The first session uses the much shorter introduction in [onboarding](references/onboarding.md), not this whole menu.
    
    ```text
    Ich bin dein Deutsch-Tutor mit Gedächtnis. Ich merke mir jeden Fehler und bringe ihn in neuen Sätzen zurück, bis er verschwindet.
    
    · Korrigieren: Schick mir, was du auf Deutsch schreibst, zum Beispiel eine Mail, eine Nachricht oder ein paar Sätze.
    · Wiederholen: „Wiederholen" oder „Lass uns üben" holt deine fälligen Fehler und Wörter.
    · Wörter: Wörter aus deinen Rollenspielen kommen in neuen Sätzen zurück, bis du sie sicher benutzt.
    · Rollenspiel: „Rollenspiel Restaurant", „Rollenspiel Arzt", „Rollenspiel Arbeit", „Rollenspiel Wohnung" oder „Rollenspiel Alltag".
    · Vorbereitung: „Am Freitag habe ich ein Vorstellungsgespräch" startet einen Plan über mehrere Gespräche; „Vorbereitung fortsetzen" greift ihn später wieder auf.
    · Sprechen: „speak restaurant" startet eine kurze Szene. Korrekturen und Wortschatz kommen erst danach.
    · Fortschritt: „Wie stehe ich?" zeigt deine DeutschDNA.
    · Einspruch: „Das war kein Fehler" nimmt eine Korrektur zurück.
    · Gemeinsam lernen: „Gib mir einen Hinweis" lässt dich selbst korrigieren. Ich merke mir, welche Hilfe funktioniert hat.
    ```
    
    ## Open with the board and the learner's own sentence
    
    Run `recap` once at the start of a conversation
  • README.mdupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

Plain-text copy, no login: /skill-bundles/ahmtsahin-deutsch-dna.md

// REPO STATS

9 stars
0 open issues
Last commit: 4d ago