The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests raw session transcripts from third-party agents which may contain untrusted data or adversarial instructions.
Ingestion points: Reads .jsonl session files from ~/.claude/, ~/.pi/, and ~/.codex/ via extract-session.js.
Boundary markers: The references/prompt-templates.md uses <session_transcript> tags, but these are for user-facing manual copy-pasting rather than programmatic enforcement.
Capability inventory: The skill uses node to execute a local parsing script and recommends file-write operations to update skills.
Sanitization: None. The script extracts content fields directly from JSON entries and formats them into a markdown transcript without escaping or filtering.
[COMMAND_EXECUTION]: The skill documentation and templates provide shell commands for the agent to execute, including script execution and file system operations.
Evidence: node "$SKILL_DIR/scripts/extract-session.js" and instructions to write improved skills back to home directory paths.

improve-skill