The Agent Skills Directory

Indirect Prompt Injection (LOW): The agent generation logic is designed to ingest and process untrusted data from local project files and user-provided descriptions to construct autonomous agent prompts.\n
Ingestion points: references/agent-creation-system-prompt.md explicitly instructs the LLM to use context from CLAUDE.md and other project files to align agent behavior.\n
Boundary markers: There are no explicit delimiters or 'ignore embedded instructions' warnings applied to the external context before it is interpolated into the generated agent's systemPrompt field.\n
Capability inventory: The examples in examples/agent-creation-prompt.md show generated agents having access to powerful file system tools such as Read, Write, Grep, and Glob.\n
Sanitization: No sanitization or escaping mechanisms are present in the generation instructions or the validation script (scripts/validate-agent.sh) to prevent malicious instructions in the source files from being promoted into the agent's core identity.\n- Command Execution (SAFE): The validation script scripts/validate-agent.sh uses standard text processing tools (grep, sed, awk, head, tail) with properly quoted variables and does not use eval or execute content from the analyzed agent files.

Agent Development