The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted design artifacts (code, screenshots, and visual parameters) into the agent's context through the 'Critic' loop described in references/critic-loop-protocol.md. This presents an indirect prompt injection surface. The skill mitigates this by using fixed, non-overridable prompt templates in templates/critic-prompt-template.md, stripping all version identifiers and metadata from candidates in 'blind' comparison modes to isolate design logic, and maintaining a 'Decision Ledger' in the design brief that prioritizes human-approved constraints over automated feedback. Sanitization is achieved through these strict boundary markers and anonymization protocols.
[COMMAND_EXECUTION]: The skill executes a local Python script, scripts/generate-seed.py, to generate random seed strings for creative exploration. This script relies on os.urandom for cryptographic entropy and performs localized base62 encoding. This is a functional requirement of the 'String Seed of Thought' methodology and does not involve untrusted network input or arbitrary command execution.
[DYNAMIC_EXECUTION]: Deterministic analysis flagged a subprocess.run call in the tests/test_generate_seed.py file. This call is used exclusively within the testing suite to validate the integrity and deterministic output of the internal seed generation script. It uses sys.executable to run its own package scripts and does not represent a runtime risk of executing external or unverified code.

taste-driven-designer