The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and assess untrusted markdown files (AGENTS.md, .claude.md) that provide instructions to AI agents. Maliciously crafted content in these files could attempt to override agent behavior during the audit phase or be suggested as project-wide improvements.\n
Ingestion points: Discovers configuration files using a shell command in SKILL.md and reads their content for scoring in references/quality-criteria.md.\n
Boundary markers: The workflow requires the agent to produce a 'Quality Report' and obtain explicit user confirmation before applying any changes using the replace_file_content tool.\n
Capability inventory: Uses find for file discovery and replace_file_content for writing updates to the filesystem.\n
Sanitization: There is no automated sanitization of external content; the skill relies entirely on the user to manually review proposed diffs for malicious instructions.\n- [COMMAND_EXECUTION]: The skill uses a shell command to locate configuration files within the current directory and references global configuration paths.\n
Evidence: find . -name "AGENTS.md" -o -name ".claude.md" -o -name ".claude.local.md" 2>/dev/null | head -50 in SKILL.md.\n- [PERSISTENCE]: The skill updates files that serve as the project's persistent 'memory' or instruction set for AI agents. By modifying AGENTS.md or global configurations in ~/.gemini/config/, changes to agent behavior can persist across multiple sessions and potentially across different projects.\n- [DATA_EXFILTRATION]: The skill identifies and accesses sensitive local configuration paths, including ~/.gemini/config/AGENTS.md and .claude.local.md. According to the skill's documentation, these files may contain 'personal/local settings' and 'user-wide defaults' that are not typically shared, leading to their exposure within the agent's context.

claude-md-improver