The Agent Skills Directory

[COMMAND_EXECUTION]: The toolkit uses a shell shim (impersonate-audit-shim.sh) and a wrapper script (impersonate-audit.sh) to manage the local environment, directories, and configuration files. These scripts automate the setup of Claude Code settings and ensure the environment is correctly prepared for an audit run.
[SAFE]: The skill includes a deterministic redactor script (scripts/redact.sh) written in Perl. The agent is explicitly instructed to pipe all log entries through this redactor before writing to disk, ensuring that Bearer tokens, JWTs, PostHog-specific prefixes (phc_, phx_, etc.), and cookies are removed from refusals.log and debug.log.
[SAFE]: A 'Refusal Protocol' is defined in SKILL.md, which provides the agent with specific logic to stop operations if the PostHog MCP plugin is scoped to the wrong project, if authentication has expired, or if the targeted experiment cannot be found. This prevents accidental data leakage or incorrect auditing.
[COMMAND_EXECUTION]: The agent is instructed to use jq and shell redirection to maintain structured logs at ~/.local/state/impersonate-audit/. This is standard behavior for Claude Code skills intended to produce durable audit trails.
[EXTERNAL_DOWNLOADS]: The installation process references npx @posthog/wizard and standard Claude plugin installation commands (/plugin install). These resources originate from the vendor's official infrastructure and are consistent with the skill's stated purpose.
[DATA_EXPOSURE]: The skill provides an optional configuration (CSM_CUSTOMER_CONTEXT_DIR) that allows the agent to read local customer notes (e.g., from an Obsidian vault). This is a user-controlled setting designed to provide context for the audit and does not constitute unauthorized access.

experiment-audit