The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Write, Edit, and Bash tools to modify configuration files such as CLAUDE.md, AGENTS.md, and creates new skill stubs (SKILL.md) in both local project directories and the user's home directory (~/.claude/). While this is the intended purpose, it represents an automated modification of the agent's operating environment.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it scans the entire conversation transcript (which may include untrusted content from the web or other files) to generate durable instructions for future sessions. A malicious payload within the processed data could attempt to trick the agent into 'learning' a rule that undermines security in future sessions.
Ingestion points: The current conversation transcript and a signal file (.aprende-signals.md) are the primary data inputs.
Boundary markers: The skill utilizes clear internal workflow steps (Pass A through Pass E) but does not apply specific sanitization to the extracted text before proposing it to the user.
Capability inventory: The agent has the ability to write executable-like skill stubs and project-level instructions using the Write and Edit tools.
Sanitization: Content is not sanitized; the system relies entirely on the human-in-the-loop confirmation step (Pass D) as the primary safeguard.
[SAFE]: The skill explicitly forbids automatic writing, requiring the user to approve specific numbered candidates before any file system changes occur. It also implements unique filename generation (appending suffixes) to prevent accidental overwriting of existing memory files.

aprende