The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from the official Prefect documentation site and reads local log files to diagnose project state.
Ingestion points: In SKILL.md, the agent is directed to fetch indices and documentation pages from docs.prefect.io. It also reads local project files like pyproject.toml and .env, and inspects CLI logs for flow runs.
Boundary markers: The skill mandates a "Verify gate" protocol where the agent must check version-sensitive claims against local evidence and explicitly cite sources, stating what remains unverified.
Capability inventory: The agent uses shell command execution (via prefect CLI) and file reading capabilities.
Sanitization: The instructions include strict requirements to filter authentication output, strip credentials and query strings from URLs, and avoid capturing unfiltered output from commands like prefect config view which might expose sensitive headers.
[COMMAND_EXECUTION]: The skill utilizes the Prefect CLI to query instance state and manage orchestration objects. It implements a "CLI-first protocol" that encourages read-only queries for auditing and requires explicit user confirmation before executing any state-changing or destructive commands (e.g., pause, cancel, delete).

prefect