mandarinat

What it does

Assists with six tasks for law education and research: (1) deep doctrinal legal research, (2) document proofreading with reference verification, Word comments, plagiarism/AI indicators, and argumentative coherence checks, (3) creating exam questions and model answers, (4) updating courses, books, and legal documents, (5) creating courses with docx and PPTX support, (6) preparing TD fiches. It triggers for any teaching/research request in law and provides guidance on when not to trigger (e.g., avoid QCM requests). It integrates tools MCP: OpenLegi, Themia, LegalDataHunter and uses Claude Code as the declared tool.

How it works

• At startup, it detects the environment (COWORK, CHAT_CU, or CHAT) to determine filesystem persistence and available tools.
• It enforces a strict five-step anti-hallucination process: Search → Find → Read → Verify backing support → Cite. It will only cite sources that have been found, read, and verified to support the asserted claim.
• It follows a detailed normative sequence for legal research: norms first, then jurisprudence (supreme courts, administrative courts, constitutional, CEDH, CJUE), then doctrine using scripts like scripts/doctrine_search.py and HAL/search, with a minimum of 10 doctrinal sources and verifiable identifiers.
• It performs a comprehensive file scan at task start, categorizes files, and leverages prior materials when available.
• It specifies task routing to five task templates (tache-1 through tache-6) and includes a playbook (tache-0) to be run before others unless a single clear legal qualification exists.
• If a user asks for a QCM, it redirects to the qcm-generator skill; otherwise it proceeds with the appropriate task sequence.
• For cassation-related analyses, it references Themia as a priority path with a three-level fallback involving OpenLegi and web_search when needed, and ensures a formal attribution link to Légifrance for cited decisions.

When to use it

Use for any teaching or research request in law: cours, TD, fiches de TD, exams, sujets, corrigés, relecture, mise à jour, doctrinal research. Do not trigger for QCM questions; use qcm-generator instead.

What it can touch

The skill declares the following tools and scripts: OpenLegi, Themia, LegalDataHunter, and Claude Code as the active coding assistant. It engages scripts such as scripts/doctrine_search.py, scripts/hal_search.py, and relies on web_search and various Légifrance/OpenLegi links as part of its workflow. It requires filesystem access in certain modes (COWORK or CHAT_CU) to manage files and run inter-session operations.

Caveats

• It enforces strict source citation: every reference cited must be found, read, and verified to support the proposition; unverified or memory-attributed references are not allowed.
• If the required tool or data source is unavailable (OpenLegi, HAL, LegalDataHunter), it falls back to web searches and informs the user of limitations.
• It explicitly excludes QCM tasks from execution and requires redirection to the appropriate skill.