The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's workflow is driven by a 'sketch plan' file, which acts as the source of implementation instructions. This creates a surface where a malicious or poorly formed plan could influence the agent's actions.
Ingestion points: The agent is instructed to read the sketch plan file path provided by the user, specifically parsing YAML frontmatter and the ## Sketch section.
Boundary markers: The skill defines a 'Non-negotiable' rule requiring explicit, affirmative user approval (e.g., 'proceed') before any implementation work occurs. It also implements a 'Chat during preparation' constraint to limit verbose output.
Capability inventory: The agent has the capability to modify application code, tests, and configuration files, and execute git add commands.
Sanitization: The instructions require the agent to reconcile the plan with the codebase and ask clarifying questions if it finds contradictions or ambiguities before proceeding.
[COMMAND_EXECUTION]: The skill automates file staging using version control tools.
Evidence: The instructions explicitly require running git add -A at the start of every fresh phase cycle.
Context: This is used as a safety mechanism to checkpoint work and ensures that the repository state is captured before new modifications are analyzed or implemented.

implement-next-phase