optimize-claude-md

Warn

Audited by Gen Agent Trust Hub on Oct 3, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_CONTEXT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_CONTEXT_INJECTION]: The skill uses the !pwd`` syntax in SKILL.md to execute the pwd command at load time. This is used to resolve the current working directory to calculate absolute paths for file operations.
  • [REMOTE_CODE_EXECUTION]: The skill invokes uvx skilllint@latest multiple times (Phases 2 and 6) to perform token measurements and completeness checks. This pattern downloads and executes a Python package from a public registry without pinning a specific version, which introduces a potential supply chain risk.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process documentation and instruction files (e.g., CLAUDE.md, SKILL.md) which are then passed to an optimization agent and a verification agent. Malicious instructions embedded in these target files could potentially influence the behavior of the agents or lead to the generation of harmful content.
  • Ingestion points: Files located at the path provided via $ARGUMENTS in SKILL.md are read and their content is passed to sub-agents.
  • Boundary markers: The skill uses XML-like tags for path resolution but lacks explicit delimiters or "ignore instructions" warnings for the content of the files being optimized when passed to the delegation agent.
  • Capability inventory: The skill has the ability to write to the filesystem (Phase 8), execute shell commands via uvx, and delegate tasks to other agents.
  • Sanitization: No sanitization or escaping is performed on the content of the target files before they are interpolated into the agent delegation prompts.
  • [COMMAND_EXECUTION]: The skill performs file system write operations in Phase 8 to overwrite the original target files with optimized content. While it requires user confirmation, the volume and complexity of automated documentation changes may make it difficult for a user to detect subtle malicious modifications introduced during the optimization process.
  • [EXTERNAL_DOWNLOADS]: The skill fetches content from openai.com for reference purposes and downloads the skilllint tool from external registries. While the reference to OpenAI is from a well-known service, the skilllint download is unversioned and originates from an untrusted source.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 3, 2026, 10:27 AM