The Agent Skills Directory

[DYNAMIC_EXECUTION]: The invocation script scripts/invoke.py dynamically locates its management engine by searching through a hardcoded list of over 60 paths in the user's home directory, targeting configuration folders for various AI agent tools (e.g., ~/.claude/skills, ~/.cursor/skills, ~/.roo/skills). The path to the executable script (kntnt.py) is computed at runtime based on these environment-specific locations.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute the discovered engine via uv run. This execution of code found through broad directory traversal poses a risk if a malicious script is placed in one of the searched locations.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from diverse project files and synthesize them into always-loaded agent instructions, creating a vulnerability to indirect prompt injection.
Ingestion points: Reads content from CLAUDE.md, AGENTS.md, agents.d/, docs/, README*, and SKILL.md files in the current repository.
Boundary markers: While the skill uses logic 'gates' to filter facts, it lacks cryptographic or clear structural delimiters to isolate untrusted project content from the instruction generation process.
Capability inventory: The skill has the ability to modify project-level configuration files (CLAUDE.md, AGENTS.md) and execute shell commands via its initialization scripts.
Sanitization: The processing logic focuses on fact utility and stability rather than sanitizing inputs for potential malicious prompt sequences.

agents-md