The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted data from external URLs (YouTube transcripts, articles, and GitHub repositories), creating a surface for instructions embedded in that content to influence the agent.
Ingestion points: External resource content is fetched and analyzed in Phase 2 using the transcribe tool and general web content extraction.
Boundary markers: The skill relies on logical instructions (Phase 5 constraints) and a "verification seam" to validate claims, but it lacks strict data delimiters or physical boundaries between untrusted content and instructions.
Capability inventory: The skill possesses significant capabilities, including local file writing, shell command execution (git, gh, yq), committing changes to the repository, and auto-merging pull requests.
Sanitization: The instructions attempt to mitigate risk by requiring credibility assessments, corroboration, and manual-style verification of all citations against the source evidence.
[COMMAND_EXECUTION]: The skill performs multiple shell operations to manage the research environment and repository state.
Evidence includes the use of yq for configuration parsing, git for branch management and updates, and gh (GitHub CLI) for creating and merging pull requests.
The shell script in Phase 0 walks up the directory tree to find configuration files, which is a standard pattern but relies on the integrity of the local file system.
[EXTERNAL_DOWNLOADS]: The skill's primary function involves fetching data from remote URLs.
It downloads video transcripts, article content, and repository metadata from the internet to perform its technical analysis.

deep-dive