The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data to generate development artifacts and code, creating a potential attack surface.
Ingestion points: Reads raw documents from the llmwiki/raw/ directory and user-provided answers in the br/interview/ folder.
Boundary markers: Uses {{...}} placeholders for variable interpolation and provides explicit scoping instructions in the revise-prompt.md template.
Capability inventory: Features file system write access to the br/ directory and executes subprocesses including python3, node, git, and the GitHub CLI.
Sanitization: Implements deterministic safety guards in loop-runner.py, including diff-jail to revert out-of-scope file changes and test-hash to detect unauthorized modifications to test suites.
[COMMAND_EXECUTION]: The skill invokes a variety of local management scripts and standard system utilities.
Evidence: Executes internal tools located in fdk/tools/ and harness/scripts/ via python3 and node. It also interacts with GitHub, a well-known service, using the gh CLI tool for issue synchronization.
[DYNAMIC_EXECUTION]: Shell commands defined within frame metadata are executed at runtime to verify task completion.
Evidence: The acceptance_test field in individual frame files contains commands that are executed by the loop-runner.py script during the code revision loop.

br