The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the web using WebSearch and WebFetch (e.g., analyzing competitor sites and design trends) to influence its design recommendations.\n
Ingestion points: WebFetch calls on competitor marketing sites and WebSearch for category trends within the 'Research Engine' and 'Design Research' phases.\n
Boundary markers: The instructions suggest summarizing findings, but do not mandate the use of technical delimiters or explicit 'ignore instructions' markers when incorporating fetched data into the prompt context.\n
Capability inventory: Access to Bash (shell), Write (file system), Agent (sub-agent spawning), and stitch*:* (UI generation).\n
Sanitization: No specific content sanitization logic is defined for the fetched external content.\n- [COMMAND_EXECUTION]: The skill defines and uses a Bash helper ss() to execute a session management script (stitch-session.mjs) using Node.js.\n
Evidence: The ss function in SKILL.md wraps calls to a local JavaScript script for state management.\n
Risk: The pattern printf '%s\\n' \"<content>\" | ss append-prd involves interpolating user-derived content directly into a Bash command string. If the LLM does not properly escape characters like backticks or dollar signs in the <content> variable, it could lead to command injection within the agent's shell environment.\n- [PERSISTENCE]: The skill implements a session state mechanism that writes data to the local file system at ~/.stitch/session/prd-draft.md to maintain state across context compactions.

stitch-ideate