The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local utility binaries located in ~/.claude/skills/gstack/bin/ (e.g., gstack-skill-start, gstack-learnings-log, gstack-skill-end) to manage session state, operational logging, and telemetry.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it ingests untrusted data from web pages.
Ingestion points: Performance metrics and page data are read from the Aside browser or a headless Chromium driver via the aside repl and $B tools.
Boundary markers: The skill defines strict behavioral rules (Rule 5) requiring the agent to treat all browser-returned content as untrusted data, not instructions. The $B tool also wraps content in explicit UNTRUSTED WEB CONTENT markers.
Capability inventory: The skill utilizes Bash for report organization and Write for saving benchmark results.
Sanitization: Data extraction is restricted to specific JSON-serialized performance metrics via in-page JavaScript execution.
[EXTERNAL_DOWNLOADS]: The skill mentions aside.com as the official source for the Aside browser. It instructs the user to download and install the application manually rather than attempting to download it programmatically.
[DYNAMIC_EXECUTION]: The preamble logic directs the agent to follow specific GSTACK_INSTRUCTION blocks for onboarding and consent workflows if they are emitted by the local gstack-skill-start binary during session initialization.

benchmark