The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted project data, such as wireframes and mockup briefs, which presents a surface for indirect prompt injection. However, the skill implements multiple robust mitigations.
Ingestion points: scripts/load_handoff.py identifies and loads paths for wireframe.md, mockup-brief.md, and design images into the agent context.
Boundary markers: Content from project-authored documents is wrapped using scripts/_untrusted_doc.py, which adds <<<UNTRUSTED_PROJECT_DOC>>> delimiters and explicit warnings to treat the content as data only. SKILL.md and reference/trust-boundaries.md provide mandatory instructions to ignore embedded commands in these assets.
Capability inventory: The skill has the capability to write Vue components and CSS files to the workspace and execute shell-based verification scripts.
Sanitization: scripts/load_handoff.py uses a whitelist for metadata keys when parsing handoff.json to prevent the injection of arbitrary free-text fields.
[COMMAND_EXECUTION]: The skill uses shell commands to run local verification and synchronization scripts.
The manifest defines tools that execute Python scripts (e.g., load_handoff.py, describe_handoff.py, verify_handoff.py) to validate implementation against specifications.
Instructions in SKILL.md and reference/subagents.md direct the agent to use standard build tools like npm test and npm run build, and to call sibling scripts within the same install tree for engineering audits.

heyeddi-handoff