The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains a 'MANDATORY' agent workflow section that directs the AI to spawn specific sub-agents (e.g., fuse-ai-pilot:explore-codebase, mcp__context7__query-docs) before and after implementation. While these appear to be vendor-specific tools related to the author 'fusengine', the use of mandatory directives to force specific tool execution is a pattern that overrides standard agent autonomy.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of untrusted user data for vector search, creating a potential surface for indirect prompt injection if the processed content contains instructions that the LLM might follow during retrieval or embedding phases.
Ingestion points: User-provided content is ingested via the VectorSearchService::ingest, VectorSearchService::search, and reembed methods as shown in references/templates/VectorSearchService.php.md.
Boundary markers: No specific delimiters or 'ignore instructions' warnings are implemented in the provided PHP templates to separate user data from system instructions.
Capability inventory: The skill utilizes database writing capabilities (Document::create, Document::update in Document-model.php.md) and makes external requests via the Laravel\Ai\Embeddings SDK to generate vectors.
Sanitization: The implementation relies on standard Laravel Eloquent protections for SQL injection, but does not explicitly sanitize the text content before it is sent to the AI embedding provider.

laravel-vector-search