The Agent Skills Directory

[PROMPT_INJECTION]: Static analysis flagged a potential instruction to conceal actions from the user. However, manual review of the 'The commit approval gate' section confirms that this is a legitimate coordination mechanism for nested skills. It specifies that if a parent skill (e.g., 'canva-implement-feedback') has already secured user approval for a specific set of changes, the editing engine does not need to ask for a redundant second confirmation. This does not bypass the human-in-the-loop requirement, as initial consent is still mandatory.
[SAFE]: The skill implements a 'Transaction Protocol' (Start -> Perform Ops -> Commit/Cancel). This is a security best practice that prevents partial or unverified changes from being applied to a user's design. All changes remain in a draft state and are lost unless the user explicitly approves the commit.
[SAFE]: The skill enforces human-in-the-loop (HITL) by requiring the agent to show the user preview thumbnails and a plain-language summary of changes. This provides a strong defense against indirect prompt injection, as any malicious modifications attempted by the agent would be visible to the user before they are finalized.
[SAFE]: The skill defines clear capability boundaries (CAN and CANNOT lists) and includes specific logic for handling 'responsive pages'. These constraints prevent the agent from attempting unsupported or potentially destructive layout operations, ensuring the skill operates only within its intended and safe scope.

canva-edit-design