The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources and the local codebase to inform its decision-making and code generation processes.
Ingestion points: External data from web searches regarding library maintenance and prior art; local codebase analysis; and user-provided brainstorming ideas (SKILL.md Phase 1).
Boundary markers: Uses a phased structure and dedicated documentation paths (e.g., docs/solutions/) to separate analysis from implementation.
Capability inventory: Arbitrary shell command execution for validation (e.g., pnpm vitest), file system operations for solution docs and experiments, and git management (git worktree).
Sanitization: Instructions mention "credential protection" during experiments, but lacks explicit sanitization or escaping protocols for data retrieved from web searches before it is incorporated into prompts or code.
[DYNAMIC_EXECUTION]: The skill involves the runtime generation and execution of code through "spikes" or experiments to verify feasibility.
Evidence: Phase 1b explicitly instructs the agent to write and run code in the project's native stack (package managers, build tools) to answer technical questions.
Mitigation: The skill mandates that these experiments be isolated in an experiments/ directory, gitignored by default, and executed in a way that protects credentials and avoids mutating live state without authorization.
[COMMAND_EXECUTION]: The skill uses various shell commands to perform development tasks and validate results.
Evidence: Instructions include running project-specific test commands (e.g., pnpm vitest), managing git worktrees for isolated changes, and using standard build tools.

dev-problem-solving