The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes task descriptions from external sources which could contain malicious instructions.
Ingestion points: Data is pulled from GitHub Issues (gh issue list), Obsidian vaults, Kanban boards, and project state files (store.json) to form 'candidate sets' for the agent to evaluate.
Boundary markers: The instructions do not define clear delimiters (such as XML tags) or safety prompts to isolate external task descriptions, allowing them to potentially influence agent behavior.
Capability inventory: The skill uses grep for file searches and audits the codebase based on ingested task parameters, providing a path for potential exploitation if malicious instructions are followed.
Sanitization: There is no evidence of input filtering or escaping for the data retrieved from external sources.
[COMMAND_EXECUTION]: The workflow relies on the execution of common developer utilities.
The agent is instructed to run grep to verify dependencies and audit the blast radius of proposed changes.
The GitHub CLI (gh) is suggested as a way to retrieve task lists from external trackers.
[EXTERNAL_DOWNLOADS]: The skill references external documentation and tools.
It links to the author's GitHub repository for methodology documentation.
It suggests using the oh-my-claudecode plugin as an environment for the multi-agent swarm.

closure-mode-at-boundaries