The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands ls and head to locate and identify the most recent transcript files within the workspace for analysis.
[INDIRECT_PROMPT_INJECTION]: The skill processes conversation transcripts containing untrusted user and tool content, presenting a surface for indirect prompt injection attacks. 1. Ingestion points: Transcript data is loaded from JSONL files as specified in the SKILL.md file discovery process. 2. Boundary markers: All reviewer and synthesizer templates in the references/ directory include explicit warnings to treat transcript data as untrusted and to ignore any embedded directives or instructions. 3. Capability inventory: The skill possesses the ability to execute shell commands, spawn parallel subagents for analysis, and perform file writes/edits to the skill repository. 4. Sanitization: The skill employs prompt-level instructions to ensure that subagents do not act upon or replicate instructions found within the analyzed data.
[SAFE]: A critical security control is the human-in-the-loop requirement in Step 5, which ensures that no skill edits or new files are applied to the repository without explicit user review and approval.

reflect