The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process *-semantic-blueprint.md files from either local storage or remote URLs. These files contain instructions and structural data that the agent interprets to draft a semantic specification. Maliciously crafted blueprints could contain indirect prompt injection intended to influence the agent's behavior during the analysis phase.
Ingestion points: Blueprints fetched via URL (using curl) or read from the local workspace in Step 0.
Boundary markers: The skill relies on structured markdown and YAML front-matter, but does not implement explicit "ignore embedded instructions" delimiters for the content of the blueprint during processing.
Capability inventory: The skill can execute shell commands via the semantius CLI (performing database operations), use curl for network requests, and execute local TypeScript scripts via bun.
Sanitization: The skill performs structural validation using a consistency-check.ts script before writing the final output, which mitigates some risks of malformed data but does not prevent logic-based prompt injection.
[EXTERNAL_DOWNLOADS]: The skill explicitly instructs the agent to use curl -s <url> to fetch blueprints from remote locations. While this is a core functional requirement, it involves fetching and processing data from potentially untrusted sources.
[COMMAND_EXECUTION]: The skill relies heavily on executing shell commands through the semantius CLI and running local helper scripts using bun (e.g., find-entities.ts, consistency-check.ts). These operations are used for legitimate model reconciliation and validation purposes.

semantius-analyst