The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs the agent to remove 'defensive framing,' 'apologetic positioning,' 'stacked hedging,' and 'obligatory cautionary endings.' While intended for academic prose refinement, these directives specifically target the rhetorical style AI models use to signal safety constraints or ethical warnings. Instructing the agent to strip such language can be used to suppress or bypass the model's built-in safety signals.
Evidence in SKILL.md: 'Remove authorial self-defense, imagined reviewer objections... empty assurances, stacked hedging, repeated scope disclaimers, and obligatory cautionary endings.'
Evidence in references/humanization-policy.md: 'Do not insert internal warnings, hidden comments, prompt notes, or workflow metadata into manuscript files.'
[COMMAND_EXECUTION]: The skill workflow includes the execution of a local Python script from a relative directory path. Running local scripts based on skill instructions requires high trust in the integrity of the surrounding project environment.
Evidence in SKILL.md: 'For full sections or papers, run ../ccf-paper-writer/scripts/check_prose_quality.py when available.'
[INDIRECT_PROMPT_INJECTION]: The skill acts as a text processor for external documents (manuscripts, designs, and experiment results) and has the capability to modify these files and execute analysis scripts. It lacks mechanisms to prevent instructions embedded within those documents from influencing the agent's behavior.
Ingestion points: Ingests prose, designs, and reported results artifacts as part of its 'Workflow' in SKILL.md.
Boundary markers: Absent; there are no instructions to ignore embedded directives in the artifacts being edited.
Capability inventory: Writing/updating files in place (ccfa-workfiles/) and executing the check_prose_quality.py script.
Sanitization: Absent; the skill does not specify any sanitization or validation of the input text before processing or script execution.

ccf-humanization