The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes conversation transcripts, which are untrusted external data sources that may contain malicious instructions meant to influence the agent or its subagents.
Ingestion points: Transcript files are read from the Claude Code project directory (typically ~/.claude/projects/) using the find-transcript.mjs script.
Boundary markers: The templates for the review subagents (references/judgment-reviewer.md, references/tooling-reviewer.md, references/divergent-reviewer.md, and references/synthesizer.md) include explicit instructions to treat the transcript as untrusted data and ignore any embedded directives or prompt injection attempts.
Capability inventory: The skill can spawn subagents using the Agent tool and eventually perform file edits on other skills, although edits are only applied after user review and approval.
Sanitization: No programmatic sanitization is performed on the transcript content; the skill relies on the model's ability to follow safety instructions.
[COMMAND_EXECUTION]: The skill executes local Node.js scripts (scripts/claude-project-dir.mjs and scripts/find-transcript.mjs) to identify and locate the current session's transcript. These scripts perform standard filesystem operations and environment variable lookups to resolve project paths.

reflect