The Agent Skills Directory

[SAFE]: The skill is implemented with a high degree of security awareness, incorporating multiple policy documents (e.g., references/constructive-roast-policy.md, references/post-run-checklist.md) and explicit consent checks to prevent abuse. The stated functionality is consistent across all provided instructions and examples.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data, such as work products and decision logs, to generate feedback. 1) Ingestion points: User requests and evidence targets identified in SKILL.md. 2) Boundary markers: Present; the skill requires an evidence ledger and strict labeling of 'receipts' (distinguishing facts from inferences). 3) Capability inventory: Low; the skill does not possess network access or high-privilege file system capabilities, utilizing only a local structure validator. 4) Sanitization: Present; the scripts/check_roast.py script validates the output structure and filters for forbidden keywords related to sensitive traits. The risk is minimized by these rigorous procedural and structural controls.
[COMMAND_EXECUTION]: The skill utilizes a local Python script (scripts/check_roast.py) for validating the structured JSON output of high-stakes roasts. A technical review of the script confirms it is a standard validator using only built-in Python libraries (argparse, json, sys, pathlib) to perform structure and content checks. It does not perform network operations, access sensitive files, or execute privileged commands.

roast-me