The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface (Category 8).
Ingestion points: Scans project source files (screens and components) to identify responsive design violations in SKILL.md Phase 2.
Boundary markers: The skill does not use specific delimiters or instructions to treat project code as untrusted data, allowing malicious instructions embedded in CSS/code comments to potentially influence agent behavior.
Capability inventory: The skill is authorized to modify source code across multiple frameworks (Flutter, React, etc.) and execute shell commands (e.g., flutter analyze, tsc, npm run test) in Phase 3 and Phase 5.
Sanitization: No escaping or validation is performed on the content read from files before it is processed by the agent.
[PROMPT_INJECTION]: The instructions contain directives for behavior override, specifically commanding the agent to operate in 'AUTONOMOUS MODE' and 'Do NOT ask questions', which suppresses user interaction and oversight during execution.
[COMMAND_EXECUTION]: Executes standard framework commands like flutter analyze, tsc --noEmit, and project test suites to verify code changes during the self-healing phase. These are legitimate development tools but represent shell command execution.
[DATA_EXFILTRATION]: Accesses the local path ~/.claude/projects/ to append execution telemetry to skill-telemetry.md. This access is used for local logging of outcomes and iterations and does not involve network exfiltration or access to sensitive credentials like SSH keys.

responsive