The Agent Skills Directory

[COMMAND_EXECUTION]: Spawns local binaries (codex and claude) to perform LLM operations. This is implemented securely using argument arrays and disabling shell execution to prevent command injection.
[COMMAND_EXECUTION]: Executes local TypeScript scripts (prepare-corpus.ts and run-agent.ts) via the Node.js runtime for workflow management.
[SAFE]: Implements comprehensive filesystem protections, including path traversal prevention through lexical and physical confinement checks within the GHOSTWRITER_HOME directory.
[SAFE]: Ingestion point: prepare-corpus.ts reads external writing samples via the --input flag. Boundary markers: buildProfilePrompt uses 'Untrusted training payload' headers. Capability inventory: spawn for CLI execution and writeFile for profile storage. Sanitization: Employs JSON.stringify to wrap samples before interpolation into prompts.
[SAFE]: Includes automated PII redaction (emails, phone numbers) and prohibits local image path inclusion to protect user privacy during the profile generation process.

train-ghostwriter