The Agent Skills Directory

[PERSISTENCE]: The skill contains logic in SKILL.md (Section 3a) that instructs the agent to 'Self-install into agent instruction files'. It forces the agent to write or replace a specific block of instructions in CLAUDE.md and AGENTS.md at the repository root. This mechanism ensures the skill's rules and TDD contract are applied in all future sessions, effectively modifying the agent's global behavior for the project.\n- [COMMAND_EXECUTION]: The evaluation script evals/run_evals.py uses subprocess.run() to execute commands via the claude CLI. Additionally, the Bash keyword provided by the skill allows for the execution of arbitrary shell commands within the test runner environment.\n- [INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant surface for indirect prompt injection. It ingests data from live web applications and external documents (PRDs, specs) in modes/discover.md, modes/interactive.md, and references/rf-recipes.md. There are no robust delimiters or sanitization routines for extracted content, and the agent has broad capabilities (shell access, file writes) that could be exploited by malicious content embedded in the pages being tested.\n- [EXTERNAL_DOWNLOADS]: The skill downloads code and configuration from several sources. It fetches its CLI from helpmetest.com and injects the axe-core library from Cloudflare's CDN (cdnjs.cloudflare.com) during accessibility audits. It also suggests downloading language runtimes like Go from go.dev. These sources are vendor-owned or well-known services.\n- [DYNAMIC_EXECUTION]: The skill uses Javascript keywords to inject and execute external scripts (e.g., axe-core) into the browser environment at runtime.

helpmetest