claude-fuer-deutsches-recht: an experimental German law skill collection with a quality lab and practice files

What the collection is, and the experiment caveat

claude-fuer-deutsches-recht is an experimental plugin and skill collection for German law, built on Anthropic's open claude-for-legal skills and fully translated into German, then adapted to the workflows of law firms, legal departments and consultants. The README is unusually direct about status: this is not a certified product but a technical playground for trying, adapting and further developing, and the collection contains no expert opinions and no legal advice, with every user expected to calibrate the skills for their own practice. The stated scope is checking documents, structured legal argumentation and producing working documents, combining subject workflows with source control, templates and separate practice files. The audience is German-speaking legal practitioners who want to see what a skill-shaped workflow looks like before building their own.

How the repository is organised

The repository is large and organised by legal area rather than by feature. The tree carries directories for employment law, employment reference checking and generation, pharmacy law, civil service law, banking legal departments, insolvency related areas and many more, plus supporting material: AGENTS.md, CLAUDE.md, CODEX.md, SKILLS.md, SCHWERPUNKTE.md, PROMPTLISTE.md, QUALITY.md, QUICKSTART.md, INSTALLATION_EINFACH.md, EVAL_RESULTS.md, TESTBERICHT.md and CHANGELOG.md, alongside two setup directories for third-party coworking setups and a EU gateway variant. The entry guidance is deliberately narrow: open a legal area, choose one entry point, provide the documents. One plugin or one of its quickstarts is usually enough, and downloading the whole collection is not a prerequisite. There is also a note warning against loading the workshop, the mini-prompt and every skill into the same chat at once, on the grounds that competing entry points produce vaguer tasks.

The Python tools and what each one generates

The Python side is small and specific to generating working documents rather than to the skills themselves. The requirements file annotates every dependency with the script that needs it.

pip install -r requirements.txt

openpyxl backs the build scripts that emit Excel files, such as liquidity planning, tax advisor tools and going-concern forecasts. PyYAML is used by scripts/sync-references.py and for parsing YAML input in the liquidity tools. python-docx converts Markdown drafts into formatted DOCX case-file documents and reads them back for the combined PDF. reportlab renders that combined PDF, pypdf appends existing PDF case documents, and odfpy renders ODT files for the single-file ZIPs. The remaining two, markdown-it-py and linkify-it-py, are pinned below their next major versions. None of this is required to use a skill; it is required to regenerate the practice files the repository ships.

The quality laboratory and the operating-modes table

Two mechanisms are worth understanding before you rely on any output. The first is the quality laboratory, described in QUALITY.md, which separates three things that are usually conflated: checking the files themselves, observing which skill actually gets selected, and evaluating the work product that comes out. The README is careful here, stating that prepared mini-prompts and prepared test cases are not model tests that have already been passed, and pointing to a process that uses actual client runs, two independent reviewers and error diagnosis, with unchecked sources and missing access left visible. Those development checks do not re-run on every matter. The second is the operating-modes table, which lists what must actually be available for each mode and what to do when it is not: for document production, if no real file tool is available, the instruction is to deliver the written text with an open export status rather than to pretend a file was created.

Source discipline and the legal-status notes

Source discipline is where this collection is most German and most careful. The README carries a legal-status note dated 10 September 2026 stating that the artificial intelligence regulation and data protection packages account for the published Digital Omnibus, the German competences under KI-MIG, and the separate data protection reform procedures that are still running, and it corrects a specific misreading: Article 4a concerns sensitive data for bias correction, not deepfakes. The current source verification row in the operating table says that with no accessible official source, you deliver the part you can substantiate and mark off the unverified statement rather than claiming clearance. The release notes for v444.5.0 apply the same discipline to case law, adding seven decisions from the BAG, BGH and BVerfG with official sources and stated limits, and noting that for one BAG proceeding only the official press release was reviewed rather than the full text, and that the release is not a complete case law overview for July and August 2026.

Limits the project states itself

The limits are stated by the project rather than discovered by the user, which is unusual and useful. It is an experiment, not a product. It offers no legal advice and no guarantee, and professional confidentiality obligations, data protection, US transfers and the AI Regulation are named as things the user remains responsible for. Different vendors implement different mechanisms for staged skill loading, document-based work orders and workflow formats, so there is no confirmed transferability of this marketplace to other systems; what matters is which functions the surface you actually use has switched on. File structure, selection budgets, prompt generation and the bundled tools are what get checked, while runtimes, automatic skill selection and file exports inside third-party user accounts are explicitly not certified. A larger model, the README adds, does not replace missing file access or an exporter.

Anthropic's original, and writing your own prompts

The obvious alternative is Anthropic's own claude-for-legal skills that this repository is based on, which is the English-language original. That is the right choice if your practice is not German, because the value here is the translation and the adaptation to German procedure, sources and official materials, not a different architecture. The other alternative is to skip skill collections entirely and write your own prompts against your own templates, which is slower at the start and gives you complete control over what goes into the context and what comes out. This repository sits between the two: it gives you a working shape to copy, and it insists you calibrate it yourself.

Licence, release cadence and version numbering

The collection is Apache-2.0 licensed, with separate LICENSE-APACHE and LICENSE-MIT files and a NOTICE at the root, plus a CODE_OF_CONDUCT.md and CONTRIBUTING.md, so the usual Apache-2.0 notice obligations apply to redistribution. Maintenance is active: the last push was on 2026-09-17, and releases come in a rapid sequence, with v444.5.2 on 2026-09-16, v444.5.0 on 2026-09-15 and v444.4.1 also on 2026-09-15. The v444.5.0 notes describe language editing across workshop entries, quickstarts and selected specialist skills, a review of the 25 existing main problem pairs, and a release check that all 26 test scripts in the repository pass. The version numbering is unusual and does not map to a semantic version, so read the changelog rather than inferring stability from the number.

Editorial conclusion

Use claude-fuer-deutsches-recht if you work in German law and want a worked example of what a skill-based workflow looks like before committing to your own, and start with a single plugin or quickstart rather than the whole collection. Do not use it as a source of legal advice or as something you can deploy unmodified, because the README states it is an experiment, offers no guarantee, and leaves selection, adaptation and substantive review with you. Verify the source discipline first: check the legal-status note dated 10 September 2026 and confirm any cited decision against its official source, since the project itself marks where only a press release was available.

Frequently asked questions

Official sources

1. Issues
2. Klotzkette/claude-fuer-deutsches-recht on GitHub
3. License: Apache-2.0
4. README
5. Releases