The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests script and scene descriptions from user-provided JSON files to automate the placement of characters on a 2D plane.
Ingestion points: The skill reads data from script.json, art.json, and outline.json in SKILL.md (Step 2 and 3) to inform the LLM about character relationships and scene geometry.
Boundary markers: Precise schemas are defined in references/schema.md, and instructional constraints are provided in references/read-pass.md and references/staging-pass.md.
Capability inventory: The agent is permitted to use Read, Write, and Bash to manipulate local project files and run the included CLI tool.
Sanitization: The renderHtml function in scripts/novel-staging.mjs uses a custom esc() function to escape HTML entities, and the JSON export mechanism in the HTML report uses Unicode escapes to prevent script injection.
[COMMAND_EXECUTION]: The workflow involves executing the local Node.js script for project seeding, data validation, and report rendering.
Evidence: SKILL.md and README.md detail commands like node scripts/novel-staging.mjs seed and node scripts/selftest.mjs.
Context: These operations are standard CLI functionalities restricted to the local environment and do not involve remote command sourcing.
[DYNAMIC_EXECUTION]: The self-test suite utilizes dynamic function construction to verify logic consistency across environments.
Evidence: scripts/selftest.mjs uses new Function() to validate the stagingCore logic that is stringified and embedded in HTML reports.
Context: This usage is confined to internal unit testing of the skill's own code and does not process external or untrusted data.

novel-staging