The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill suggests the installation of the @colbymchenry/codegraph package from the npm registry and the use of various Docker images (including lizard, jscpd, dependency-cruiser, golangci-lint, actionlint, helm, sqlfluff, and pghero). All suggested downloads and images are pinned to specific versions or SHA256 digests to ensure supply chain integrity. Furthermore, no software is installed automatically; all such actions are presented as suggestions within a report that requires explicit user approval through a natural-language gate.
[COMMAND_EXECUTION]: The skill includes scripts (scripts/lab-probe.sh and scripts/lab-probe.ps1) designed to detect host environment capacity (RAM, CPU, disk) for optional Docker-based analysis. It also suggests running one-shot, read-only Docker containers to gather evidence for its critiques. These commands are transparently defined and are only executed after the user approves the exact command string provided in the output gate.
[INDIRECT_PROMPT_INJECTION]: Because the skill's primary function is to analyze untrusted data (user-provided plans, code, and diffs), it has a large attack surface for indirect prompt injection. The skill mitigates this risk through a comprehensive 'Safety boundaries' section (SKILL.md §6), which explicitly instructs the agent that all analyzed content must be treated as data, not instructions, and that any directives embedded within the analyzed material must be disregarded. The mandatory evidence-based reporting and the final user-approval gate provide additional layers of human-in-the-loop protection.

devils-advocate