The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted external data, including existing site content, project briefs, and wiki context, to generate site concepts and prototypes. This design represents a surface for indirect prompt injection attacks where instructions embedded in the input data could potentially influence the agent's behavior.
Ingestion points: reports/semantic-core.json, docs/site-structure.md, project briefs, and wiki context (referenced in references/concept-prototype-contract.md).
Boundary markers: None identified.
Capability inventory: File system writes to docs/ and reports/ directories, and execution of local validation scripts.
Sanitization: No evidence of input sanitization or filtering was found.
[COMMAND_EXECUTION]: The SKILL.md file defines a validation workflow that executes local Python scripts (quick_validate.py and lint_production_skill.py) located within the $CODEX_HOME environment path. These commands are used for linting and validating the skill's integrity and do not appear to process direct user input or perform network operations.

concept-prototyper