The Agent Skills Directory

[PROMPT_INJECTION]: Potential for indirect prompt injection. The skill extracts and displays summaries and messages from past Claude Code sessions. If a past session contains malicious instructions, the agent processing the output of this skill might inadvertently follow them.
Ingestion points: scripts/dig.py reads session history from .jsonl files in ~/.claude/projects/.
Boundary markers: The rendering instructions in SKILL.md do not include delimiters or warnings to ignore embedded instructions.
Capability inventory: The skill performs read-only operations using ls and python3.
Sanitization: The script truncates session summaries and message snippets to 80 characters.
[COMMAND_EXECUTION]: The skill executes the ls command to discover project directories and the ghq list -p command within the Python script to resolve repository paths.
[DATA_EXFILTRATION]: The skill accesses sensitive interaction logs in ~/.claude/projects/, which contain private conversation history and metadata from past sessions.

dig