The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple subcommands, including the core autoresearch loop, fix.md, and regression.md, execute shell commands provided as Verify, Guard, or Target arguments. These commands are executed within an autonomous loop. While a 'Safety Screen' is mentioned in autoresearch.md to block dangerous patterns like rm -rf or curl|bash, this protection relies entirely on the AI agent's adherence to the instruction rather than a hard sandbox or technical restriction.
[PROMPT_INJECTION]: The skill processes untrusted data from multiple ingestion points without hard boundaries:
improve.md performs web searches to gather insights which then influence the generation of Product Requirement Documents (PRDs).
learn.md, predict.md, and security.md scan the entire codebase and its history to build a knowledge base used for decision-making.
plan.md and probe.md take natural language goals to derive execution parameters and shell commands.
Sanitization: The skill relies on natural language instructions for safety screening rather than robust input validation or structured data handling.
Capability Inventory: The agent possesses extensive capabilities, including executing shell commands, performing git operations (commit, revert, worktree), and writing files to the local system.
[EXTERNAL_DOWNLOADS]: The improve.md subcommand performs WebSearch operations to synthesize research findings. This represents a continuous ingestion of external content into the agent's context during its iteration loop.
[DATA_EXFILTRATION]: The improve.md subcommand utilizes external searches to research product improvements. There is a risk that sensitive information extracted from the codebase (such as internal project names, architecture details, or configuration keys) could be included in search queries directed to third-party search providers.

autoresearch