The Agent Skills Directory

[COMMAND_EXECUTION]: Interacts with a local CLI tool reflect located at ~/.claude/bin/reflect to list, mark, and update the status of learnings.
[DATA_EXFILTRATION]: Reads local code snippets (before and after blocks) and learning descriptions from a local ledger and transmits them to GitHub issues. The skill explicitly includes logic to genericise these snippets (removing private identifiers and path names) when filing to public repositories or global scopes.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the reflect ledger (specifically the text, before, and after fields which may originate from code comments or logs). This poses a surface for indirect prompt injection where malicious instructions in a code comment could influence the agent's routing or merging logic.
Ingestion points: Step 1 loads pending learnings via reflect marker ls (SKILL.md).
Boundary markers: Uses a proposal table to prompt the user before filing.
Capability inventory: Performs reflect marker mark subprocess calls and network-based issue filing via file-issue (SKILL.md).
Sanitization: Instructions include genericising public-repo issues to remove private identifiers (SKILL.md). This risk is mitigated by the requirement for explicit user acceptance before filing.

triage-reflections