The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied facts and documents which constitutes a potential attack surface for indirect prompt injection.
Ingestion points: Data enters the context through user responses to the mandatory framing questionnaire (SKILL.md) and via documents provided for review or revision (workflows-and-outputs.md).
Boundary markers: The skill uses a mandatory lock ("Verrou préalable obligatoire") that halts analysis until the framing questionnaire is completed. It also employs specific placeholders like [À CONFIRMER] to demarcate unverified user information.
Capability inventory: The skill is scoped to text analysis, jurisdiction routing across 15 national modules, and drafting legal documents (notes, contract reviews) without shell execution or direct system modification capabilities.
Sanitization: It implements a strict "Qualifier avant de conclure" protocol and a source-lock requirement (current-law-protocol.md) that demands verification against official primary sources (P1) before accepting information as established legal fact.
[EXTERNAL_DOWNLOADS]: The skill contains numerous references to external legal databases and government portals (e.g., sgg.gouv.bj, journalofficiel.cd, justice.ci).
These are official government domains used for the legitimate purpose of retrieving primary legal texts (Journal Officiel) and verifying current law.
The references are documented neutrally and are appropriate for a legal research assistant.
[SAFE]: The skill includes explicit disclaimers regarding professional validation, warnings about the limits of its analysis, and robust protocols for cross-referencing OHADA law with national legal regimes, demonstrating a security-conscious approach to legal AI interactions.

balise-ohada