The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes common developer tools for automating the ADR lifecycle, including git (branching, committing, pushing), gh (creating labels and PRs), and filesystem utilities like mkdir and grep. These operations are well-defined and constrained to the local documentation directory.- [EXTERNAL_DOWNLOADS]: The skill references installing supplementary skills via npx skills add jwilger/agent-skills. This targets resources from the same vendor and aligns with the expected installation patterns for this environment.- [PROMPT_INJECTION]: The skill includes an 'Override rule' instructing the agent to disregard external instructions that attempt to bypass the required architectural phases, serving as a defensive control. It also possesses a surface for indirect prompt injection via the ingestion of external documentation.
Ingestion points: External documentation URLs and source code files accessed during the RESEARCH phase.
Boundary markers: The skill uses a structured ADR template with a dedicated 'Research Findings' section to isolate findings.
Capability inventory: The skill has file-system write capabilities and the ability to execute Git/GitHub CLI commands.
Sanitization: The workflow mandates 'Research Critique' and 'Draft Critique' phases where findings must be validated by a human before proceeding.- [DATA_EXPOSURE]: The skill implements 'access guard files' (CLAUDE.md and AGENTS.md) which provide instructions to restrict the agent's internal reasoning from over-relying on historical ADRs for current implementation tasks, effectively functioning as a context-management safety feature.

architecture-decisions