The Agent Skills Directory

[SAFE]: The skill implements a comprehensive and disciplined development lifecycle (Research, Grill, Plan, Stage, Review, Commit) that emphasizes security and structure. It includes mandatory instructions to redact secrets in both documentation (references/document.md) and context handoff files (references/handoff.md). The research process (references/research.md) is designed to run in a subagent when supported, providing a layer of isolation for exploratory tasks. Verification of code changes relies on the repository's native test and lint tooling, ensuring that the agent operates within the project's established safety boundaries.
[INDIRECT_PROMPT_INJECTION]: The skill by design reads and processes external data from the repository and chat history, which constitutes a surface for indirect prompt injection.
Ingestion points: Research involves reading repository files and exploration guides (references/research.md); the review phase processes git diffs and code symbols (references/review.md); documentation tasks utilize chat history and project artifacts (references/document.md).
Boundary markers: The skill employs structured artifacts (plans, handoffs) and explicit directives like "Redact secrets" and "Preserve unrelated user changes" to help distinguish between instructions and data.
Capability inventory: The skill can perform file writes (restricted primarily to .agentflow/ and docs/), execute git commands for diffing and history inspection, and run project-defined verification scripts (references/review.md).
Sanitization: The skill delegates safety to the agent's core capabilities and the project's own AGENTS.md guidelines for handling shell command arguments and data interpolation.

agentflow