The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze conversation history which includes untrusted data such as raw user input and external tool outputs.
Ingestion points: Conversation history is ingested in SKILL.md (Process Step 1) and passed to reviewer sub-agents.
Boundary markers: Robust security warnings are included in all reviewer and synthesizer prompts (references/divergent-reviewer.md, references/judgment-reviewer.md, references/tooling-reviewer.md, references/synthesizer.md) instructing the agent to treat history as untrusted and ignore any embedded instructions.
Capability inventory: The skill has the ability to modify repository files (via skill-creator) and write to external issue trackers, though these actions are mediated by specialized tools.
Sanitization: A mandatory human-in-the-loop approval step (Step 5) is implemented for all state-changing edits, ensuring the user reviews and authorizes any skill modifications before they are applied.
[EXTERNAL_DOWNLOADS]: Sub-agents are authorized to use MCP tools to retrieve data from external services to verify evidence.
The prompts in references/*.md permit the use of MCP tools to look up referenced context from sources like ticket trackers, documentation, and observability platforms. These lookups are restricted to read-only access and are strictly scoped to evidence cited within the task history.

reflect