The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill captures potentially untrusted content from the current conversation and codifies it into a permanent instruction file. If an attacker can influence the conversation to include malicious instructions that appear successful, the agent may save them as a reusable skill.
Ingestion points: The agent is instructed to "Walk back through the conversation" to distill steps (SKILL.md).
Boundary markers: The skill requires the agent to quote the understood workflow back to the user for confirmation before writing (Step 1).
Capability inventory: The skill utilizes file writing (mkdir -p, write to SKILL.md) and reading (cat / read) capabilities.
Sanitization: The instructions explicitly direct the agent to "Strip secrets and PII" and replace sensitive data with placeholders (Step 3 and Edge Cases).
[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands to create directories and write Markdown files to the local file system at ~/.smtm/projects/{slug}/skills/{name}/SKILL.md.
[PERSISTENCE]: The skill establishes a persistence mechanism by saving instruction sets to a local directory that is designed to be "auto-loaded" and surfaced in every future session within the same project. This allows instructions to persist across agent restarts and new sessions.
[DYNAMIC_EXECUTION]: The skill dynamically generates a new SKILL.md file from conversation data and a predefined template. These generated files contain instructions that the agent is expected to execute or follow in subsequent interactions.

money-skillify