The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data generated during task execution to update durable documentation sinks, which could be leveraged if the processed data contains malicious instructions.
Ingestion points: The skill reads from active runtime phase sheets (found at state/.../phases/phase-NN.md) and summarizes descriptions of work produced and lessons learned.
Boundary markers: The instructions do not specify any delimiters, escaping, or safety markers to differentiate between the agent's instructions and the data being harvested.
Capability inventory: The skill modifies local markdown files (notes.md) and executes several commands from the doctrine CLI toolset.
Sanitization: There is no mention of sanitization or validation of the harvested content before it is written to the durable documentation sinks.
[COMMAND_EXECUTION]: The skill explicitly instructs the agent to execute specific CLI tools to manage and verify project state.
Evidence: Direct instructions to use doctrine slice list to identify slices, doctrine slice notes <ID> to scaffold notes, and doctrine check gate to verify code modifications.

harvest