The Agent Skills Directory

[PROMPT_INJECTION]: The skill describes patterns where external data (e.g., task.md, spec.md, SHARED_TASK_NOTES.md) is ingested and interpolated directly into agent prompts using shell commands like $(cat file). This creates a surface for indirect prompt injection if those files contain malicious instructions.
Ingestion points: SKILL.md provides templates using cat to read task descriptions and specifications into the prompt context.
Boundary markers: The provided script snippets lack explicit delimiters or instructions to ignore embedded commands within the ingested content.
Capability inventory: The loops have significant capabilities including file system writes, shell command execution (via bun/npm test), and Git operations.
Sanitization: No sanitization or validation of the ingested file content is performed in the example implementation patterns.
[COMMAND_EXECUTION]: The implementation examples utilize shell scripts to orchestrate the workflow, executing CLI tools such as claude, git, and test runners (bun, npm). While these are standard for developer-centric autonomous workflows, they represent a significant capability that should be executed within a sandboxed environment.

autonomous-loops