The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill relies on a suite of external subagents (stack-researcher, blueprint-writer, blueprint-validator) and specialized tools (ui-ux-pro-max, emil-design-eng, agent-browser, browser-harness, claude-api) that are dispatched dynamically during the architectural process.
[DATA_EXFILTRATION]: The architecture process involves accessing external websites, including those 'behind a login' using the browser-harness tool. This introduces potential risks regarding the handling and exposure of user credentials or session tokens during site analysis.
[COMMAND_EXECUTION]: Rule 7 requires the generation of 'verify commands' (commands that exit 0) for each build step in the final blueprint. This means the skill generates shell commands intended for later execution by the user or another agent, creating a vector for command injection if the design phase is compromised.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests large amounts of untrusted data from external reference sites and existing codebases (in 'BROWNFIELD' mode) to inform its architectural decisions.
Ingestion points: User-provided reference URLs, existing local code repositories, and live software registries.
Boundary markers: The skill does not define specific boundary markers for external data, though it mandates a confirmation gate before final generation.
Capability inventory: Includes file system writes (to ./blueprints/), network requests (via WebSearch and agent-browser), and the generation of executable shell commands.
Sanitization: Employs a blueprint-validator subagent to audit the generated output for compliance with internal rules before delivery to the user.

architect