The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external design briefs and evidence to construct prompts for a secondary designer agent.
Ingestion points: The agent reads designs/brief.md and results from the Mobbin MCP search_screens tool (references/designer-iteration.md).
Boundary markers: The skill uses 'HARD RULE' instructions to constrain the designer sub-agent, but does not employ strict delimiters to isolate untrusted input from the brief or evidence.
Capability inventory: The skill utilizes shell execution (pen, git, cp, open, sips), file system writes, and MCP tool calls (SKILL.md, references/designer-iteration.md).
Sanitization: There are no explicit steps to sanitize or validate the content of the design brief or Mobbin search results before interpolation.
[DYNAMIC_EXECUTION]: The skill requires the agent to generate and execute local scripts based on provided logic.
The agent is instructed to create snapshot-tree.sh and diff-tree.sh in designs/scripts/ to perform JSON tree hashing and comparison (references/gated-runs.md).
Python and JavaScript snippets are provided for token migration and semantic auditing of the .pen JSON files (references/token-migration.md).
[COMMAND_EXECUTION]: The skill relies on various shell commands for its core functionality.
Interacts with the pen CLI for design tasks and pen interactive for flushing app changes.
Uses standard utilities like git, cp, open, sips (for image cropping), magick/compare (for pixel diffing), and awk (SKILL.md, references/gated-runs.md).
[EXTERNAL_DOWNLOADS]: The skill references external resources and installation of additional tools.
Recommends the installation of the higgsfield-generate skill from higgsfield-ai/skills (references/designer-iteration.md).
Fetches design evidence and screenshot comparisons from external sources like Mobbin and Google Fonts (via the Export tool).

pen-dev