The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The self-improving-agent plugin contains features that ingest data from potentially untrusted sources, such as shell tool outputs and project memory files, creating a potential surface for indirect prompt injection.
Ingestion points: The error-capture.sh hook in self-improving-agent/hooks/ processes CLAUDE_TOOL_OUTPUT, and the memory-analyst.md agent reads MEMORY.md and associated topic files from ~/.claude/projects/.
Boundary markers: The error capture hook wraps ingested data in <error-detected> tags, but does not provide explicit instructions to the model to ignore embedded commands. The memory analyst agent does not use specific delimiters for its input files.
Capability inventory: The agent environment includes capabilities for shell command execution, file system modification, and network access, which could be targeted by a successful injection.
Sanitization: No robust sanitization or safety filtering is performed on the ingested content beyond basic truncation in the error hook.
[EXTERNAL_DOWNLOADS]: The documentation and installation guides recommend using external tools to download and execute code from unverified sources.
Evidence: README.md suggests installation via npx ai-agent-skills install alirezarezvani/claude-skills/engineering-team.
Details: These tools download and execute code from a repository managed by an individual user who is not part of a verified trusted organization, introducing a potential supply chain risk.

engineering-skills