The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data to generate launch assets and checklists, creating a surface for indirect prompt injection.
Ingestion points: The skill gathers product context from user-provided sources such as product docs and positioning notes (referenced in SKILL.md).
Boundary markers: The skill architecture enforces a human-in-the-loop pattern where all agent-generated drafts must be reviewed and approved in the #/assets UI route before execution.
Capability inventory: Through scripts/execute_decisions.mjs, the skill can mark items as approved for hand-off to other specialized skills (like kelly-email) for public submission or sending.
Sanitization: The skill relies on manual human review and approval as the primary sanitization mechanism for generated content.
[COMMAND_EXECUTION]: The skill includes several utility scripts that perform file system and environment operations.
Evidence: scripts/execute_decisions.mjs, scripts/setup.mjs, and scripts/publish_airapp.mjs execute logic that reads from environment variables (BUSABASE_API_KEY, BUSABASE_BASE_URL) and interacts with the Busabase API.
[EXTERNAL_DOWNLOADS]: The skill depends on several external Node.js packages for its web interface and data management.
Evidence: package.json and content/kelly-launch-app/package.json list dependencies such as busabase-sdk, hono, and @hono/node-server.
[DYNAMIC_EXECUTION]: The development environment uses a build-time tool for bundling code.
Evidence: The build:sdk script in content/kelly-launch-app/package.json uses esbuild-wasm to bundle the busabase-sdk into the application's vendor directory.

kelly-launch