The Agent Skills Directory

[COMMAND_EXECUTION]: The skill operates primarily via the circuit CLI, executing commands such as circuit run, circuit resume, and circuit preview. These commands manage coding flows, audit progress, and interact with AI connectors. The instructions provide specific guidance on shell safety, recommending that user-provided goal text be single-quoted and apostrophes escaped to prevent unintended shell expansion or command injection.
[DYNAMIC_EXECUTION]: The skill supports the dynamic generation and execution of task flows. The circuit generate command uses an LLM to propose flow logic, which is then subjected to offline validity and runnability checks before execution. Additionally, the skill allows for custom connectors, which are user-defined executables that can be assigned to specific research or review roles. These are documented as trusted local processes.
[INDIRECT_PROMPT_INJECTION]: As the skill is designed to process user-defined task descriptions and project source code, it possesses an attack surface for indirect prompt injection. The skill mitigates this by instructing the agent to maintain strict shell boundaries and sanitizing user inputs.
Ingestion points: User-provided task descriptions are ingested through the --goal and --why flags in circuit run commands.
Boundary markers: The agent is directed to use single quotes around the goal text to prevent the expansion of variables or command substitution.
Capability inventory: The circuit CLI can read and write files (in Fix and Build flows) and invoke external worker processes via configured connectors.
Sanitization: The skill's instructions explicitly require the agent to escape embedded apostrophes as '\'' to ensure the integrity of the shell command.

circuit