The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell scripts (scripts/run-webgpt-review.sh) which in turn invoke other tool-providing scripts (ask/run.sh and surf/run.sh) located in sibling directories. These executions use bash arrays to safely handle arguments and prevent shell injection.
[EXTERNAL_DEPENDENCIES]: The skill relies on the presence of specific sibling skills (ask, surf, and best-practices-skills) within the environment's directory structure for core functionality and automated testing.
[INDIRECT_PROMPT_INJECTION]: The skill implements an attack surface for indirect prompt injection by design, as its primary purpose is to ingest external data (via the --bundle file) and user-provided --instructions for processing by a downstream agent (WebGPT).
Ingestion points: External bundle files (concatenated text or zip) and user-supplied instruction strings.
Boundary markers: The skill contract specifies that bundles must be browser-readable artifacts, but does not implement explicit delimiters to separate data from instructions within the bundle.
Capability inventory: The skill has the ability to execute subprocesses and interface with browser automation tools.
Sanitization: The script performs basic validation on the bundle's file type and existence but does not sanitize the contents of the file before submission.

webgpt-review