The Agent Skills Directory

[PROMPT_INJECTION]: The instructions mandate the agent to "Treat the skill file as executable instructions, not reference," which is a directive aimed at overriding default safety, planning, or reasoning protocols.
[DYNAMIC_EXECUTION]: The preamble section uses high-risk patterns such as eval "$(~/.claude/skills/gstack/bin/gstack-slug)" and source <(~/.claude/skills/gstack/bin/gstack-repo-mode) to execute shell logic generated dynamically by local binaries at runtime.
[PERSISTENCE]: The skill contains logic to proactively modify the project's CLAUDE.md file by injecting a "Skill routing" section. This ensures that the agent will continue to invoke this specific skill suite for a wide range of future tasks, effectively establishing long-term persistence within the project's development workflow.
[REMOTE_CODE_EXECUTION]: The skill implements an "Inline upgrade flow" triggered by gstack-update-check. If an upgrade is available, the skill is instructed to download and follow new instructions from a remote source, which constitutes a self-updating remote code execution vector.
[DATA_EXFILTRATION]: An integrated telemetry system collects metadata including repository names, branch names, session IDs, and execution durations. This data is logged locally and, if opted-in, transmitted externally via the gstack-telemetry-log binary.
[COMMAND_EXECUTION]: The skill executes a significant volume of shell commands and internal binaries during its preamble to inspect environment variables, session history, and git configurations before the primary task begins.

document-generate