The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill involves a multi-step loop where an AI subagent generates Python rendering code which is then executed at runtime to produce artifacts.
Evidence: In phases/03-implement.md, the '@analysis-implementer' subagent is tasked with writing a renderer script to a specific directory and then executing it via python3.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from a user-configurable directory and reads configuration from project files, creating an attack surface where malicious data could influence agent behavior.
Ingestion points: The skill reads research datasets from a directory resolved in phases/01-precheck.md and configuration from the project's AGENTS.md file.
Boundary markers: The instructions in phases/03-implement.md explicitly command the subagent to 'Read input files from inside this directory only' and 'Do not read from anywhere else on disk'.
Capability inventory: The skill possesses the capability to execute shell commands via bash and perform file system writes (PDF, PNG, and JSON metadata).
Sanitization: There is no programmatic sanitization of input data; the skill relies on prompt-based instructions to the AI agents to maintain safety boundaries.
[COMMAND_EXECUTION]: The skill uses the shell to execute generated code and verify the existence of resulting artifacts.
Evidence: phases/03-implement.md contains instructions to 'Run it via Bash: python3 <code_dir>/.py' and use ls -la to check for the output file.

figure-bake