The Agent Skills Directory

[DATA_EXFILTRATION]: The skill records detailed logs of every AI interaction, including prompt templates and full outputs, to workspace/memory/prompts/prompt-log.jsonl. This behavior exposes potentially sensitive user data, PII, or hardcoded logic in plaintext on the local filesystem.
[PROMPT_INJECTION]: The update prompt command allows for the manual overriding of any skill's prompt template. This mechanism could be abused to replace legitimate skill logic with malicious instructions if the command is triggered by an attacker.
[PROMPT_INJECTION]: The skill implements an automated 'Prompt Evolution' feature that modifies and promotes prompt templates based on historical outcomes and user feedback. This design is susceptible to indirect prompt injection (Category 8), as malicious inputs could trick the optimizer into adopting harmful instructions as the primary prompt.
Ingestion points: Reads historical interaction data from prompt-log.jsonl and user feedback strings to generate variations.
Boundary markers: No delimiters or instructions are used to separate original prompt logic from untrusted external feedback during the evolution process.
Capability inventory: Utilizes file:read and file:write permissions to modify log files and update the primary instructions of other skills.
Sanitization: No validation or sanitization is performed on the user feedback or AI-generated prompt variations before they are accepted as 'winners'.

prompt-optimizer