The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates and executes shell scripts (evaluate.sh) within its workspace to facilitate automated benchmarking and quality scoring. This is a core functional requirement of the autonomous improvement loop.
[EXTERNAL_DOWNLOADS]: The documentation references an implementation harness located at https://github.com/a-tokyo/agent-skills-harness. As this repository belongs to the skill's author, it is considered a vendor-owned resource.
[PROMPT_INJECTION]: The skill ingests arbitrary files from a user-specified input/ directory to guide the synthesis of new instructions, creating an indirect prompt injection surface. Adversarial content in these files could theoretically influence the output of the subagent pipeline.
Ingestion points: The orchestrator and researcher subagents are instructed to scan and read all materials in the input/ folder during Phases 1 and 2.
Boundary markers: The provided prompts for research subagents do not explicitly mandate the use of delimiters or 'ignore' instructions for processing external content.
Capability inventory: The factory can write to the local filesystem and execute scripts that perform network operations via curl to LLM endpoints.
Sanitization: The skill relies on the synthesis and verification agents to judge the quality of content without an explicit pre-processing step to sanitize untrusted input for malicious commands.

create-skill-autoresearch