The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill manages a workflow for other agent skills, creating a surface where instructions embedded in the files being processed could influence the agent's operations.
Ingestion points: The agent reads and interprets SKILL.md and evals.json files during the drafting, polishing, and verification phases.
Boundary markers: There are no explicit markers or instructions to ignore embedded prompts in the data being processed, though evaluation subagents are given a restricted context.
Capability inventory: The skill can write to the file system, execute local scripts, and dispatch subagents with specific capabilities for git and deployment tasks.
Sanitization: The workflow does not explicitly sanitize instructions from the skills being edited before they are followed by evaluation executors.
[COMMAND_EXECUTION]: The skill automates the execution of local shell and Python scripts as part of its testing and synchronization logic.
Evidence: Calls to evals/files/init.sh and a referenced gen_fixtures.py are used to set up evaluation environments.
Evidence: The synchronization step (sync.sh) is triggered automatically during the 'Ship' phase based on project-specific rules defined in the repository.
[DYNAMIC_EXECUTION]: The skill dynamically generates prompt contexts for subagents to perform specialized evaluation and deployment tasks.
Evidence: Evaluation executors are dispatched using templates that incorporate the skill-under-test directly into the subagent's system-level instructions.
Evidence: The deployment phase delegates sensitive operations to a subagent using a dynamically constructed handoff containing project paths and post-ship commands.

write-great-skill