The Agent Skills Directory

[PERSISTENCE]: The skill instructs the agent to modify global, user-level instruction and configuration files for multiple platforms, including ~/.claude/CLAUDE.md, ~/.cursor/rules/pstack-models.mdc, ~/.config/goose/.goosehints, ~/.config/amp/AGENTS.md, and others. By adding persistent instructions to these files, the skill ensures its logic and model mappings are loaded in every future agent session, establishing long-term persistence within the agent's environment.
[DATA_EXPOSURE]: The skill defines operations to locate and read agent session transcripts, event logs, and interaction databases across various harnesses (e.g., ~/.claude/projects/, ~/.copilot/session-state/, sessions.db, crush.db). These files contain history of previous interactions, which may include sensitive information, code, or credentials discussed in other sessions.
[INDIRECT_PROMPT_INJECTION]: The skill ingests session transcripts through the find-transcripts operation. These transcripts contain data from previous turns and potentially external sources. The skill lacks instructions for sanitizing or delimiting this ingested content, creating a vulnerability where instructions embedded in previous interaction data could influence the agent's current or future actions.
[COMMAND_EXECUTION]: The skill mapping includes instructions for shell execution and plugin installation across different environments, such as amp -ox "<prompt>", goose plugin install <git-url>, and ln -sfn "$PWD"/skills/* ~/.agents/skills/. These provide a surface for executing system commands or installing external packages.

harness