The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines subagents that process external data, creating an attack surface where malicious instructions could be embedded in ingested content.\n
Ingestion points: sonnet-researcher.md (web pages/URLs), fable-reviewer.md (external specifications and artifacts), and sonnet-collector.md (files and web pages).\n
Boundary markers: Present. SKILL.md instructs the lead model to treat subagent output as data to weigh rather than instructions, and fable-reviewer.md defines its input as external.\n
Capability inventory: Subagents such as opus-builder and fable-builder have implementation capabilities including file system modifications, while sonnet-researcher has network and file access.\n
Sanitization: The skill relies on model judgment rather than explicit technical sanitization for processed data.\n- [COMMAND_EXECUTION]: The skill's subagent definitions and harness adapters involve shell command execution. Subagents like opus-builder are tasked with running validation checks defined in packets, and the skill suggests a logging configuration using jq to process tool input and append it to a local log file.\n- [EXTERNAL_DOWNLOADS]: The skill fetches technical guidance and configuration details from Anthropic's official platforms and Cursor's documentation site. It also references academic research papers from ArXiv and community routing examples.

fable5-model-routing