The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes JSON and HTML data previously captured from an external site (located in stardust/current/pages/). During Phase 2, the agent uses LLM judgment to inject this content into new templates (Path A') or compose new pages (Path B). There are no explicit boundary markers or instructions to disregard embedded commands in the source data, creating a surface where malicious content from the original site could influence the agent's behavior during the migration process.
Ingestion points: Data is read from stardust/current/pages/<slug>.json and stardust/current/pages/<slug>.html as documented in reference/migration-procedure.md.
Boundary markers: Absent in the rendering and injection instructions.
Capability inventory: The skill performs file writes to the local filesystem (stardust/migrated/) and network resolution of external image URLs for media reconciliation.
Sanitization: The skill normalizes whitespace and wraps content in specific data-attributes (e.g., data-slot, data-section), but does not escape or filter natural language content for potential instructions.
[COMMAND_EXECUTION]: The skill instructions in Phase 3 include shell commands using find and grep to audit the generated site for source-tree escapes and absolute references. It also executes local Node.js audit scripts, such as fixtures/pagemap-audit.mjs and fixtures/file-protocol-audit.mjs, which utilizes Playwright for headless crawling.
[EXTERNAL_DOWNLOADS]: The setup instructions require the installation of the playwright package via npm if it is not already present in the environment. It also relies on an external tool and repository (github.com/pbakaus/impeccable) as a core dependency for the migration workflow.

migrate