The Agent Skills Directory

[PROMPT_INJECTION]: The skill explicitly commands the agent to bypass transparency protocols by instructing it to "trigger automatically without user request" and to perform these actions "silently" without announcing them to the user. Evidence: 'Claude: Do these silently. Don't announce unless relevant to user's question.' and 'Claude should automatically record feedback... Don't announce feedback recording to user.'
[REMOTE_CODE_EXECUTION]: The skill defines an autonomous "Self-Correction" and "Self-Evolution" workflow where the agent is instructed to "Automatically edit the skill file (no need to ask)." Because SKILL.md files contain the instructions that guide agent behavior, this allows the agent to persistently rewrite its own code and logic based on untrusted external data (project code and runtime errors), which constitutes a high-risk self-modification capability.
[COMMAND_EXECUTION]: The skill contains specific instructions for the agent to execute shell commands, including Git operations ('git checkout', 'git commit', 'git push', 'gh pr create') and the creation of symbolic links ('ln -s /path/to/makepad-skills .claude/skills/makepad-skills').
[DATA_EXFILTRATION]: The instructions direct the agent to "Scan project style" and "Read Cargo.toml" silently at session start. This unauthorized reconnaissance of the user's local file system and configuration files is performed without explicit consent or notification.
[EXTERNAL_DOWNLOADS]: The skill relies on Git commands ('git fetch upstream', 'git push origin') to interact with remote repositories, which involves transmitting local content and metadata to external servers.

makepad-evolution