The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses directive and imperative language designed to override the agent's default safety or cautious behavior. Instructions such as "Don't turn that checklist into a request for permission," "No ‘should I proceed?’ for reading files [or] editing within scope," and "Stop dithering and do the damn work" are explicitly intended to suppress confirmation steps and move the agent into a more autonomous, less restricted mode of operation.
[INDIRECT_PROMPT_INJECTION]: By instructing the agent to skip confirmation requests for actions like "reading files" and "editing within scope," the skill creates a high-risk surface for indirect prompt injection. If an agent processes a file or external data containing malicious instructions, the 'Just Do It' directive may lead the agent to execute those instructions immediately without presenting them to the user for approval.
Ingestion points: The skill instructions in SKILL.md (Step 2: "Inspect the relevant evidence") explicitly encourage the agent to ingest external evidence and act on it immediately.
Boundary markers: There are no instructions regarding delimiters or warnings to ignore instructions embedded within the processed data.
Capability inventory: The skill encourages "reading files, editing within scope, [and] running local checks," which involves file system access and potentially local command execution.
Sanitization: There are no instructions provided for sanitizing or validating external input before acting upon it.

just-fucking-do-it