The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input to generate workflow artifacts, creating a surface for indirect prompt injection.
Ingestion points: User-provided intentions are stored in the .itw/<id>/intake file.
Boundary markers: The content of the intake file is interpolated into various phase prompts (e.g., clarification.md, prd.md, issues.md) using the {{INITIAL_INTENTION}} template token.
Capability inventory: The skill manages local files and executes shell commands via the itw CLI. It also defines future execution agents (Worker, Reviewer) with capabilities like workspace-write and tools such as git and rg.
Sanitization: No explicit sanitization or escaping of the user-provided intention is performed before it is interpolated into templates.
[COMMAND_EXECUTION]: The skill executes shell commands for installation and self-verification.
Evidence: src/intent_to_workflow/core.py uses subprocess.run to execute the itw command to verify its installation fingerprint.
Evidence: The scripts/setup and scripts/setup.ps1 scripts use uv tool install to install the package globally.

intent-to-workflow