The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's automated review process involves subagents generating Python scripts to apply findings and edits to markdown artifacts.
Evidence: MERGE-BRIEF.md instructs the subagent to 'Write (spec-review-edits.py or plan-review-edits.py) beside the artifact... Run it once.'
Evidence: PLAN-REVIEW-BRIEF.md and SPEC-REVIEW-BRIEF.md also contain instructions for subagents to generate and execute local Python scripts to modify the project's plan and spec files.
[COMMAND_EXECUTION]: The instructions direct the agent to execute various shell commands discovered within the project environment or defined during the planning phase.
Evidence: PLAN.md mandates the agent to 'captured and run once' every 'Gates: command' (e.g., test, lint, build) and probe 'every toolchain behavior' such as build tools and bundlers.
Evidence: PLAN-REVIEW-BRIEF.md instructs reviewers to 'run it yourself' when verifying 'unprobed oracles' or third-party tool outputs.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input (raw ideas) and codebase content, which is then processed by multiple subagents with broad file-system and command-execution capabilities.
Ingestion points: SKILL.md (initial ask), BRAINSTORM.md (raw ideas), and SPEC.md (problem descriptions).
Boundary markers: Subagent instructions are delimited by tags, but there are no explicit instructions to ignore potentially malicious directions embedded within the ingested data.
Capability inventory: The skill has the ability to write arbitrary files, generate/execute Python code, and run shell commands.
Sanitization: There is no evidence of sanitization or strict schema validation for the data interpolated into subagent prompts.

crank