The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external software and data-analysis requests, which inherently creates an indirect prompt injection surface.
Ingestion points: User-provided requests are processed in SKILL.md and through the flows described in the reference files.
Boundary markers: The skill establishes clear 'Human review boundaries' and a 'Verification Contract' in references/data-analysis.md to mitigate the risk of the agent executing or planning unsafe instructions.
Capability inventory: Capabilities include creating plan files and recommending the execution of shell commands for Git operations and code tangling.
Sanitization: The instructions explicitly mandate: 'Keep secrets and raw sensitive records out of plans, tickets, and logs.'
[DYNAMIC_EXECUTION]: The skill facilitates a workflow where code is generated from documentation and executed.
Evidence: references/data-analysis.md describes a process where 'tangled Python' is generated from Org-mode source blocks and subsequently executed or tested, which represents planned dynamic execution of generated content.
[COMMAND_EXECUTION]: The planning framework involves the use of shell commands for repository integrity and reproducibility.
Evidence: The skill instructs the implementation agent to capture git rev-parse HEAD and provides placeholders for commands to tangle and run analyses within the generated plans.

grill-to-plan