The Agent Skills Directory

[PROMPT_INJECTION]: The skill's primary function is to process untrusted data from the conversation history and work sessions to generate new instructions (skills). This creates an indirect prompt injection surface where a malicious input in the work session could be captured into a new skill. However, this is mitigated by the skill's design, which includes high confidence ROI checks (Decision Framework) and a mandatory human-in-the-loop verification (Step 5) where the draft is presented for user review before finalization. Ingestion points include any content monitored during the work session, while templates in 'references/skill-templates.md' serve as structural boundaries. The capability inventory is limited to writing markdown files.
[COMMAND_EXECUTION]: The skill instructs the agent to save draft skill files to a local directory (/mnt/user-data/outputs/). This involves legitimate file system operations consistent with the skill's stated purpose of persistent knowledge capture and does not involve the execution of arbitrary shell commands.

learning-capture