The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Bash to execute a local script (scripts/audit-context) that performs broad file system discovery using commands like find, ls, and grep to analyze context usage across the environment.
[DATA_EXFILTRATION]: The skill reads sensitive internal application data, including conversation history files (~/.claude/projects/*/*.jsonl), plugin configurations, and MCP server settings (~/.claude/settings.json). This access is used to calculate token usage and identify context composition targets.
[INDIRECT_PROMPT_INJECTION]: The skill's auditing process ingests data from untrusted sources, including the content of other skills, project-specific memory files (~/.claude/projects/*/memory/*.md), and session logs, which could contain malicious instructions designed to influence the audit report.
Ingestion points: scripts/audit-context reads various files across the ~/.claude/ directory and project-specific directories to build its inventory.
Boundary markers: The audit report generates tables and summaries; no explicit instruction-boundary markers (such as XML tags or specific 'ignore instructions' warnings for the content of the audited files) are implemented in the reporting logic.
Capability inventory: The skill has access to Bash, Read, Grep, and Glob, enabling extensive system inspection and metadata extraction.
Sanitization: The script extracts data points like file sizes and token counts, but metadata such as tool arguments extracted from session logs are not explicitly escaped before being presented to the agent.

context-audit