The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from the target codebase, including documentation and architectural decision records, which serves as an attack surface for indirect prompt injection.
Ingestion points: Reads existing documentation such as ARCHITECTURE.md, CONTEXT.md, and decision records in docs/adr/ to understand the system's current state.
Boundary markers: The instructions do not specify explicit delimiters or "ignore embedded instructions" markers for these ingestion points.
Capability inventory: The skill can perform file-write operations to update architectural documentation (ARCHITECTURE.md, CONTEXT.md) and generate new specifications (docs/specs/). It also suggests running project-specific CLI tools (e.g., uv, cargo, pnpm, prisma).
Sanitization: No explicit sanitization or validation of the ingested documentation content is described before processing or interpolating into sub-agent prompts.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for verifying package versions and metadata using official, well-known registries and service domains.
Evidence: tooling/PYTHON.md fetches metadata from pypi.org; tooling/RUST.md references the official rust-lang.org distribution manifest; tooling/TYPESCRIPT.md references registry.npmjs.org for version checks.
[DYNAMIC_EXECUTION]: The skill employs a parallel multi-agent pattern to explore alternative interface designs for architectural components.
Evidence: INTERFACE.md describes the process of spawning three or more parallel sub-agents, each provided with distinct technical briefs to generate varied implementation strategies.
[COMMAND_EXECUTION]: The skill instructions and tooling guides mention common development commands for environment management, linting, and database migrations.
Evidence: Recommends usage of uv run, cargo test, pnpm vitest, and npx prisma migrate for standard development workflows.

architecture