The Agent Skills Directory

[DYNAMIC_EXECUTION]: The self-improving-agent skill implements a workflow to promote patterns from MEMORY.md to CLAUDE.md and .claude/rules/. This effectively allows the agent to dynamically generate and apply its own governing instructions based on runtime observations.
[COMMAND_EXECUTION]: Multiple skills, including ms365-tenant-manager, aws-solution-architect, and azure-cloud-architect, provide templates for high-privilege operations using PowerShell and shell commands. These include user provisioning, security policy modifications, and infrastructure deployment.
[PROMPT_INJECTION]: The adversarial-reviewer skill utilizes persona framing (e.g., 'Sabotador', 'Auditor de Segurança') to force the agent into specific behavioral modes for code auditing. While functional for its stated purpose, it employs role-play techniques often associated with prompt manipulation.
[INDIRECT_PROMPT_INJECTION]: The self-improving-agent skill represents an indirect injection surface. It ingests data from MEMORY.md (which can contain logs, error messages, and external content) and uses that data to influence future agent behavior via rule promotion in CLAUDE.md. There are no explicit sanitization or boundary markers defined in the markdown instructions to prevent adversarial content in the memory from being promoted to permanent rules.
[EXTERNAL_DOWNLOADS]: The google-workspace-cli skill instructs the agent to install external tools via npm install -g @anthropic/gws and cargo install gws-cli. The @anthropic NPM scope is associated with a trusted organization. The skill also fetches resources from GitHub releases and common CDNs (e.g., jsDelivr for GSAP).
[DATA_EXPOSURE]: The snowflake-development skill provides instructions for accessing sensitive credentials via os.environ["SNOWFLAKE_PASSWORD"]. This is documented as a best practice for secret management (environment variables) rather than an unsafe exposure.

engineering-skills