The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest untrusted data from project logs and memory files and promote them to high-priority instruction files, creating a significant injection surface.
Ingestion points: The agent reads from MEMORY.md, project-specific topic files, and captures shell command error outputs at runtime.
Boundary markers: No specific delimiters or "ignore instructions" warnings are defined to prevent the agent from executing instructions potentially embedded within the captured logs or memory entries.
Capability inventory: The skill possesses the capability to write to CLAUDE.md and .claude/rules/*.md (which govern agent behavior) and generate new standalone skill files (SKILL.md).
Sanitization: There is no evidence of content validation, escaping, or filtering before promotion, allowing potentially malicious instructions in project data to become persistent project rules.
[EXTERNAL_DOWNLOADS]: The skill provides installation instructions for several third-party sources not identified as trusted vendors.
The installation guide references a GitHub marketplace repository (alirezarezvani/claude-skills) and an external platform (clawhub.ai).
It suggests running a local shell script ./scripts/codex-install.sh for skill deployment, which executes arbitrary local code.
[COMMAND_EXECUTION]: The skill includes a hook mechanism (error-capture) that actively monitors and captures output from Bash commands.
The PostToolUse hook captures shell output after every command to extract error details, which are then fed into the memory system.

self-improving-agent