The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that could lead the agent to bypass standard safety practices.
Evidence: The "Hard rules" section explicitly instructs the agent to "Skip unnecessary abstractions/tests/safeguards," which may result in the reduction of security-relevant checks during implementation.
Indirect Prompt Injection Surface: The skill processes external Markdown plan files to drive its automated task loop.
Ingestion points: The agent reads a user-provided "plan" file (SKILL.md).
Boundary markers: No specific delimiters or instructions to ignore embedded commands are present when the plan is ingested.
Capability inventory: The skill utilizes subagents, browser automation (agent-browser), file write access, and shell execution (lint/build tools).
Sanitization: No validation or sanitization of the plan content is performed before implementation.
[COMMAND_EXECUTION]: The task loop involves executing shell-based development tools to verify code changes.
Evidence: The "Check" phase includes running lint, typecheck, build, and migration tools, as well as runtime cleanup commands like Pi clean.
[EXTERNAL_DOWNLOADS]: The skill utilizes tools for fetching information from the web to guide its implementation.
Evidence: The "Analyze" and "Check" steps utilize web-research and agent-browser for gathering information on third-party behaviors and validating UI changes.

implement-plan