romainsimon/paperasse

What it solves

Paperasse provides a set of specialized skills to help AI agents navigate the complex bureaucracy of French accounting, taxation, invoicing, notary services, and corporate auditing. It transforms general-purpose AI agents into expert copilots capable of handling specific French administrative tasks, reducing errors in tax filings and financial reporting.

How it works

The project consists of a collection of Markdown-based skills that can be integrated into various AI agents (such as Claude Code, Cursor, or Windsurf). These skills provide the agent with the ground-truth knowledge of French laws (CGI, BOFiP, NEP), forms, and deadlines.

To enhance functionality, it includes:

• Connectors: Automatic retrieval of bank transactions (Qonto) and payments (Stripe).
• Scripts and Templates: Node.js scripts for deterministic calculations (amortization, VAT) and templates for generating professional PDFs, financial statements, and the FEC (Fichier des Écritures Comptables).
• Evaluation Framework: An automated testing suite that measures the accuracy of the agent's responses with and without the skills using an LLM-as-judge approach.

Who it’s for

It is designed for business owners, accountants, and individuals dealing with the French administrative system who use AI agents to automate their bookkeeping, tax preparation, and legal documentation.

Highlights

• Six Specialized Roles: Includes dedicated skills for accountants, tax inspectors, statutory auditors, personal tax experts, notaries, and co-ownership managers.
• CGI & BOFiP Integration: Grounded in official French tax and accounting standards.
• End-to-End Workflow: Supports a full annual closing cycle from transaction classification to tax simulation.
• Deterministic Tools: Uses JS scripts for precise financial calculations to avoid LLM hallucinations.
• Automated Evals: Built-in benchmarking to prove the performance delta provided by the skills.