The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes conversation transcripts and uses the content to generate instructions that are permanently added to agent definition files. This provides a vector for indirect prompt injection if malicious instructions are embedded within the analyzed conversation history.
Ingestion points: Conversation transcripts are read from the file path provided to scripts/signal_detector.py or through the PreCompact hook input.
Boundary markers: The skill does not implement delimiters or 'ignore' warnings for the data it interpolates into agent instruction files.
Capability inventory: The skill possesses the ability to write to agent instruction files (such as ~/.claude/agents/), create new skill definitions (.claude/skills/), and execute Git commands for versioning changes.
Sanitization: The scripts/signal_detector.py uses basic regex to extract signals but does not sanitize the resulting instruction strings for potential injection patterns before proposing them for persistence.
[DYNAMIC_EXECUTION]: The skill dynamically generates and saves new SKILL.md files based on content extracted from conversations. These files are subsequently loaded as instructions by the agent platform in future sessions. The generation logic relies on templates located in references/skill_template.md.

reflect-learn