French Legal Drafting & Review
An Agent Skill for drafting and reviewing legal documents, built for the French and EU context — designed to know when the available facts aren't enough to draw a conclusion.
AI-generated legal text can be fluent and wrong at the same time.
Invented facts
Confident-sounding legal text can be produced even when the underlying facts were never actually provided.
Over-interpretation
A quick legal qualification can turn a possibility into a stated certainty.
Rules applied without context
Generic legal language doesn't always fit a specific product, audience or data flow.
Legal rules vs. platform rules
App store, marketplace or platform requirements get treated as legal obligations — or the reverse.
Not just generating the right text.
The skill's main goal isn't simply to produce legal text. It enforces a methodology meant to avoid invented facts, hasty qualifications, categorical conclusions when information is missing, and confusion between legal obligations, platform rules, and GDPR roles.
"Legal AI is not only about generating the right text. It is also about knowing when the available facts are not enough to generate it."
- Traceability of assumptions
- Explicit flagging of missing information
- Caution in legal qualification
- Clear separation between the publishable document and internal analysis
From the facts on the table to the delivered document.
Seven principles against invention and over-reach.
Fact-first
Never invent missing facts.
Qualification before conclusion
Determine the legal qualification before drawing conclusions.
Explicit uncertainty
When the available facts are insufficient, say so explicitly.
Treatment-by-treatment GDPR analysis
Do not automatically classify every service provider or every dataset the same way.
Legal rules ≠ platform rules
Clearly distinguish French/EU legal obligations from Apple, Google, marketplace or other platform requirements.
Consistency checking
Look for contradictions between clauses, documents and assumptions.
Currentness
Flag rules and information that require verification at the date of publication.
Documents built for the French and EU context.
Stress-tested against the failure modes it's meant to avoid.
These scenarios probe, among other things:
- Factual invention
- Legal over-interpretation
- Contradictions
- Missing information
- Controller / processor GDPR distinction
- Platform rules
- Potentially outdated information
These are internal project tests — not a legal certification and not an independent audit.
Three worked cases in the repository.
Open source, MIT-licensed
Released under the MIT license as redaction-legale-numerique.skill, the distributable package of the skill.
Explore the skill on GitHub.
Source code, documentation and examples are all in the repository.
View on GitHub