claude-context-improver
Warn
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: MEDIUMDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill is designed to interact with and modify files in the ~/.claude directory, including the global CLAUDE.md and rules files. This directory contains configuration that influences the agent across all projects, making it a sensitive file path. Evidence: The Phase 1 discovery process explicitly uses find commands to locate and read CLAUDE.md and rule files within the $HOME/.claude directory.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests content from repository-level CLAUDE.md and rule files to assess their quality and apply improvements. This creates a surface where malicious or malformed instructions in the repository could be processed by the agent to influence its behavior or suggested changes. 1. Ingestion points: Discovery phase (Phase 1) searches for CLAUDE.md and .claude/rules/*.md files in the repository. 2. Boundary markers: The workflow includes a mandatory Quality Report (Phase 4) for user review before any edits are performed. 3. Capability inventory: The skill utilizes Write, Edit, and restricted Bash tools to modify the repository context. 4. Sanitization: A structured rubric-based assessment (Phase 3) is used to evaluate the content against specific context-engineering principles before generating proposals.
Audit Metadata