The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon untrusted external data, which creates a surface for indirect prompt injection attacks.
Ingestion points: Processes task statements, pre-analysis results, grilling outputs, Figma mockups, and repository-specific standards files (e.g., CLAUDE.md, .claude/standards/).
Boundary markers: The instructions emphasize 'Surgical changes only' and include a mandatory handover report to keep the human developer informed. It also defines 'Stop rules' to halt execution if contradictions or review failures occur.
Capability inventory: The skill can perform git commits, execute shell commands (tests, directory listings), and initiate network requests (API calls for verification).
Sanitization: Includes a mandatory self-review loop using a built-in code-review skill and a specific security-review pass for sensitive changes involving authentication or secrets.
[COMMAND_EXECUTION]: The skill uses local shell commands for project exploration and verification.
Evidence: Utilizes ls -L, find -L, and grep to locate and navigate standards in Step 1. It also executes project-specific test suites and potentially docker compose or curl commands in Step 4 for environment verification.
[EXTERNAL_DOWNLOADS]: The skill connects to external services and environments to perform its tasks.
Evidence: Accesses Figma via the figma-design-to-code skill and official Figma MCP to retrieve design structure and screenshots. It also performs network requests (API calls, Playwright UI testing) against defined 'stand' URLs for live verification.
[DYNAMIC_EXECUTION]: The skill generates and executes content at runtime based on repository discovery.
Evidence: Dynamically generates the .claude/verify.md configuration file if missing, based on discovered project infrastructure, and subsequently executes verification steps defined within that file.

implement-task